Skip to contentSkip to conversation

fio mmap randread 64k more than 40% regression with 2.6.33-rc1

PATCH Zhang, Yanmin 5 participants 17 messages Dec 31, 2009 – Jan 20, 2010 Block layer

lore
Share & tools
lore
#0 lore raw

Comparing with kernel 2.6.32, fio mmap randread 64k has more than 40% regression with 2.6.33-rc1.

The test scenario: 1 JBOD has 12 disks and every disk has 2 partitions. Create 8 1-GB files per partition and start 8 processes to do rand read on the 8 files per partitions. There are 8*24 processes totally. randread block size is 64K.

We found the regression on 2 machines. One machine has 8GB memory and the other has 6GB.

Bisect is very unstable. The related patches are many instead of just one.

1) commit 8e550632cccae34e265cb066691945515eaa7fb5

Author: Corrado Zoccolo <[email protected]>
Date:   Thu Nov 26 10:02:58 2009 +0100

cfq-iosched: fix corner cases in idling logic

This patch introduces about less than 20% regression. I just reverted below section and this part regression disappear. It shows this regression is stable and not impacted by other patches.

@@ -1253,9 +1254,9 @@ static void cfq_arm_slice_timer(struct cfq_data *cfqd)

                return;

        /*
-        * still requests with the driver, don't idle
+        * still active requests from this queue, don't idle
         */
-       if (rq_in_driver(cfqd))
+       if (cfqq->dispatched)
                return;

2) How about other 20%~30% regressions? It's complicated. My bisect plus Li Shaohua's investigation located 3 patches, df5fe3e8e13883f58dc97489076bbcc150789a21, b3b6d0408c953524f979468562e7e210d8634150, 5db5d64277bf390056b1a87d0bb288c8b8553f96.

tiobench also has regression and Li Shaohua located the same patches. See link
http://lkml.indiana.edu/hypermail/linux/kernel/0912.2/03355.html.

Shaohua worked about patches to fix the tiobench regression. However, his patch doesn't work for fio randread 64k regression.
I retried bisect manually and eventually located below patch,

commit 718eee0579b802aabe3bafacf09d0a9b0830f1dd

Author: Corrado Zoccolo <[email protected]>
Date:   Mon Oct 26 22:45:29 2009 +0100

cfq-iosched: fairness for sync no-idle queues

The patch is a little big. After many try, I found below section is the key. @@ -2218,13 +2352,10 @@ cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,

        enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);

        if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
-           (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
+           (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))
                enable_idle = 0;

That section deletes the condition checking of !cfqd->cfq_latency, so enable_idle=0 with more possibility.

I wrote a testing patch which just overlooks the original 3 patches related to tiobench regression, and a patch which adds back the checking of !cfqd->cfq_latency. Then, all regression of fio randread 64k disappears.

Then, instead of working around the original 3 patches, I applied Shaohua's 2 patches and added the checking of !cfqd->cfq_latency while also reverting the patch mentioned in 1). But the result still has more than 20% regression. So Shaohua's patches couldn't improve fio rand read 64k regression.

fio_mmap_randread_4k has about 10% improvement instead of regression. I checked that my patch plus the debugging patch have no impact on this improvement.

randwrite 64k has about 25% regression. My method also restores its performance.

I worked out a patch to add the checking of !cfqd->cfq_latency back in function cfq_update_idle_window.

In addition, as for item 1), could we just revert the section in cfq_arm_slice_timer?

As Shaohua's patches don't work for this regression, we might continue to find better methods. I will check it next week.

---

With kernel 2.6.33-rc1, fio rand read 64k has more than 40% regression. Located below patch.

commit 718eee0579b802aabe3bafacf09d0a9b0830f1dd

Author: Corrado Zoccolo <[email protected]>
Date:   Mon Oct 26 22:45:29 2009 +0100

cfq-iosched: fairness for sync no-idle queues

It introduces for more than 20% regression. The reason is function cfq_update_idle_window forgets to check cfqd->cfq_latency, so enable_idle=0 with more possibility.

Below patch against 2.6.33-rc1 adds the checking back.

---

block/cfq-iosched.c +2 -2
diff -Nraup linux-2.6.33_rc1/block/cfq-iosched.c linux-2.6.33_rc1_rand64k/block/cfq-iosched.c
@@ -3064,8 +3064,8 @@ cfq_update_idle_window(struct cfq_data *
 		cfq_mark_cfqq_deep(cfqq);
 
 	if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
-	    (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
-	     && CFQQ_SEEKY(cfqq)))
+	    (!cfqd->cfq_latency && !cfq_cfqq_deep(cfqq) &&
+		sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))
 		enable_idle = 0;
 	else if (sample_valid(cic->ttime_samples)) {
 		if (cic->ttime_mean > cfqd->cfq_slice_idle)
#1 lore raw

Hi Yanmin,

Quoted 2 lines · Yanmin

Comparing with kernel 2.6.32, fio mmap randread 64k has more than 40% regression with 2.6.33-rc1.

Can you compare the performance also with 2.6.31?
I think I understand what causes your problem.
2.6.32, with default settings, handled even random readers as sequential ones to provide fairness. This has benefits on single disks and JBODs, but causes harm on raids.
For 2.6.33, we changed the way in which this is handled, restoring the enable_idle = 0 for seeky queues as it was in 2.6.31:
@@ -2218,13 +2352,10 @@ cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,

       enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);

       if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
-           (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
+           (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))
               enable_idle = 0;
(compare with 2.6.31:
        if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
            (cfqd->hw_tag && CIC_SEEKY(cic)))
                enable_idle = 0;

excluding the sample_valid check, it should be equivalent for you (I assume you have NCQ disks))
and we provide fairness for them by servicing all seeky queues together, and then idling before switching to other ones.

The mmap 64k randreader will have a large seek_mean, resulting in being marked seeky, but will send 16 * 4k sequential requests one after the other, so alternating between those seeky queues will cause harm.

I'm working on a new way to compute seekiness of queues, that should fix your issue, correctly identifying those queues as non-seeky (for me, a queue should be considered seeky only if it submits more than 1 seeky requests for 8 sequential ones).

Quoted 23 lines

The test scenario: 1 JBOD has 12 disks and every disk has 2 partitions. Create 8 1-GB files per partition and start 8 processes to do rand read on the 8 files per partitions. There are 8*24 processes totally. randread block size is 64K.

We found the regression on 2 machines. One machine has 8GB memory and the other has 6GB.

Bisect is very unstable. The related patches are many instead of just one.

1) commit 8e550632cccae34e265cb066691945515eaa7fb5
Author: Corrado Zoccolo <[email protected]>
Date:   Thu Nov 26 10:02:58 2009 +0100

   cfq-iosched: fix corner cases in idling logic

This patch introduces about less than 20% regression. I just reverted below section and this part regression disappear. It shows this regression is stable and not impacted by other patches.

Review static void cfq_arm_slice_timer(struct cfq_data *cfqd) · line 1261

Quoted 10 lines
(quoted diff) +2 -2
@@ -1253,9 +1254,9 @@ static void cfq_arm_slice_timer(struct cfq_data *cfqd)
                return;
 
        /*
-        * still requests with the driver, don't idle
+        * still active requests from this queue, don't idle
         */
-       if (rq_in_driver(cfqd))
+       if (cfqq->dispatched)
                return;
 

This shouldn't affect you if all queues are marked as idle. Does just your patch:

block/cfq-iosched.c cfq_update_idle_window(struct cfq_data * · line 3068

Quoted 4 lines
block/cfq-iosched.c +2 -2
-           (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
-            && CFQQ_SEEKY(cfqq)))
+           (!cfqd->cfq_latency && !cfq_cfqq_deep(cfqq) &&
+               sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))

fix most of the regression without touching arm_slice_timer?

I guess

Quoted 1 line

5db5d64277bf390056b1a87d0bb288c8b8553f96.

will still introduce a 10% regression, but this is needed to improve latency, and you can just disable low_latency to avoid it.

Thanks,
Corrado

#2 lore raw
Quoted 6 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines · Yanmin

Comparing with kernel 2.6.32, fio mmap randread 64k has more than 40% regression with 2.6.33-rc1.

Thanks for your timely reply. Some comments inlined below.

Quoted 1 line

Can you compare the performance also with 2.6.31?

We did. We run Linux kernel Performance Tracking project and run many benchmarks when a RC kernel is released.

The result of 2.6.31 is quite similar to the one of 2.6.32. But the one of 2.6.30 is about 8% better than the one of 2.6.31.

Quoted 4 lines

I think I understand what causes your problem.
2.6.32, with default settings, handled even random readers as sequential ones to provide fairness. This has benefits on single disks and JBODs, but causes harm on raids.

I didn't test RAID as that machine with hardware RAID HBA is crashed now. But if we turn on hardware RAID in HBA, mostly we use noop io scheduler.

Quoted 11 lines

For 2.6.33, we changed the way in which this is handled, restoring the enable_idle = 0 for seeky queues as it was in 2.6.31:

(quoted diff) +1 -1
@@ -2218,13 +2352,10 @@ cfq_update_idle_window(struct cfq_data *cfqd,
 struct cfq_queue *cfqq,
       enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
 
       if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
-           (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
+           (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))
               enable_idle = 0;

(compare with 2.6.31:

Quoted 7 lines
        if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
            (cfqd->hw_tag && CIC_SEEKY(cic)))
                enable_idle = 0;

excluding the sample_valid check, it should be equivalent for you (I assume you have NCQ disks))
and we provide fairness for them by servicing all seeky queues together, and then idling before switching to other ones.

As for function cfq_update_idle_window, you is right. But since 2.6.32, CFQ merges many patches and the patches have impact on each other.

Quoted 36 lines

The mmap 64k randreader will have a large seek_mean, resulting in being marked seeky, but will send 16 * 4k sequential requests one after the other, so alternating between those seeky queues will cause harm.

I'm working on a new way to compute seekiness of queues, that should fix your issue, correctly identifying those queues as non-seeky (for me, a queue should be considered seeky only if it submits more than 1 seeky requests for 8 sequential ones).

Quoted 25 lines

The test scenario: 1 JBOD has 12 disks and every disk has 2 partitions. Create 8 1-GB files per partition and start 8 processes to do rand read on the 8 files per partitions. There are 8*24 processes totally. randread block size is 64K.

We found the regression on 2 machines. One machine has 8GB memory and the other has 6GB.

Bisect is very unstable. The related patches are many instead of just one.

1) commit 8e550632cccae34e265cb066691945515eaa7fb5

Author: Corrado Zoccolo <[email protected]>
Date:   Thu Nov 26 10:02:58 2009 +0100

cfq-iosched: fix corner cases in idling logic

This patch introduces about less than 20% regression. I just reverted below section and this part regression disappear. It shows this regression is stable and not impacted by other patches.

(quoted diff) +0 -0
@@ -1253,9 +1254,9 @@ static void cfq_arm_slice_timer(struct cfq_data *cfqd)
               return;
 
Quoted 7 lines

Review static void cfq_arm_slice_timer(struct cfq_data *cfqd) · line 1260

Quoted 7 lines
(quoted diff) +2 -2
       /*
-        * still requests with the driver, don't idle
+        * still active requests from this queue, don't idle
        */
-       if (rq_in_driver(cfqd))
+       if (cfqq->dispatched)
               return;

Although 5 patches are related to the regression, above line is quite independent. Reverting above line could always improve the result for about 20%.

Quoted 2 lines
Quoted 1 line

This shouldn't affect you if all queues are marked as idle.

Do you mean to use command ionice to mark it as idle class? I didn't try it.

Quoted 7 lines

Does just
your patch:

Review

Quoted 4 lines
(quoted diff) +2 -2
-           (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
-            && CFQQ_SEEKY(cfqq)))
+           (!cfqd->cfq_latency && !cfq_cfqq_deep(cfqq) &&
+               sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))

fix most of the regression without touching arm_slice_timer?

No. If to fix the regression completely, I need apply above patch plus a debug patch. The debug patch is to just work around the 3 patches report by Shaohua's tiobench regression report. Without the debug patch, the regression isn't resolved.
Below is the debug patch.

block/cfq-iosched.c +8 -1
diff -Nraup linux-2.6.33_rc1/block/cfq-iosched.c linux-2.6.33_rc1_randread64k/block/cfq-iosched.c
@@ -592,6 +592,9 @@ cfq_set_prio_slice(struct cfq_data *cfqd
 	cfqq->slice_start = jiffies;
 	cfqq->slice_end = jiffies + slice;
 	cfqq->allocated_slice = slice;
+/*YMZHANG*/
+	cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
+
 	cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
 }
 
@@ -1836,7 +1839,8 @@ static void cfq_arm_slice_timer(struct c
 	/*
 	 * still active requests from this queue, don't idle
 	 */
-	if (cfqq->dispatched)
+	//if (cfqq->dispatched)
+	if (rq_in_driver(cfqd))
 		return;
 
 	/*
@@ -1941,6 +1945,9 @@ static void cfq_setup_merge(struct cfq_q
 		new_cfqq = __cfqq;
 	}
 
+	/* YMZHANG debug */
+	return;
+
 	process_refs = cfqq_process_refs(cfqq);
 	/*
 	 * If the process for the cfqq has gone away, there is no
Quoted 5 lines

I guess

Quoted 1 line

5db5d64277bf390056b1a87d0bb288c8b8553f96.

will still introduce a 10% regression, but this is needed to improve latency, and you can just disable low_latency to avoid it.

You are right. I did a quick testing. If my patch + revert 2 patches and keep 5db5d64, the regression is about 20%.

But low_latency=0 doesn't work like what we imagined. If patch + revert 2 patches and keep 5db5d64 while set low_latency=0, the regression is still there. One reason is my patch doesn't work when low_latency=0.

Quoted 3 lines

Thanks,
Corrado

I attach the fio job file for your reference.

I got a cold and will continue to work on it next week.

Yanmin

[global]
direct=0
ioengine=mmap
size=8G
bs=64k
numjobs=1
loops=5
runtime=300
group_reporting
invalidate=0
directory=/mnt/stp/fiodata
file_service_type=random:36

[job_sdb1_sub0]
startdelay=0
rw=randread filename=data0/f1:data0/f2:data0/f3:data0/f4:data0/f5:data0/f6:data0/f7:data0/f8

[job_sdb1_sub1]
startdelay=0
rw=randread filename=data0/f2:data0/f3:data0/f4:data0/f5:data0/f6:data0/f7:data0/f8:data0/f1

[job_sdb1_sub2]
startdelay=0
rw=randread filename=data0/f3:data0/f4:data0/f5:data0/f6:data0/f7:data0/f8:data0/f1:data0/f2

[job_sdb1_sub3]
startdelay=0
rw=randread filename=data0/f4:data0/f5:data0/f6:data0/f7:data0/f8:data0/f1:data0/f2:data0/f3

[job_sdb1_sub4]
startdelay=0
rw=randread filename=data0/f5:data0/f6:data0/f7:data0/f8:data0/f1:data0/f2:data0/f3:data0/f4

[job_sdb1_sub5]
startdelay=0
rw=randread filename=data0/f6:data0/f7:data0/f8:data0/f1:data0/f2:data0/f3:data0/f4:data0/f5

[job_sdb1_sub6]
startdelay=0
rw=randread filename=data0/f7:data0/f8:data0/f1:data0/f2:data0/f3:data0/f4:data0/f5:data0/f6

[job_sdb1_sub7]
startdelay=0
rw=randread filename=data0/f8:data0/f1:data0/f2:data0/f3:data0/f4:data0/f5:data0/f6:data0/f7

[job_sdb2_sub0]
startdelay=0
rw=randread filename=data1/f1:data1/f2:data1/f3:data1/f4:data1/f5:data1/f6:data1/f7:data1/f8

[job_sdb2_sub1]
startdelay=0
rw=randread filename=data1/f2:data1/f3:data1/f4:data1/f5:data1/f6:data1/f7:data1/f8:data1/f1

[job_sdb2_sub2]
startdelay=0
rw=randread filename=data1/f3:data1/f4:data1/f5:data1/f6:data1/f7:data1/f8:data1/f1:data1/f2

[job_sdb2_sub3]
startdelay=0
rw=randread filename=data1/f4:data1/f5:data1/f6:data1/f7:data1/f8:data1/f1:data1/f2:data1/f3

[job_sdb2_sub4]
startdelay=0
rw=randread filename=data1/f5:data1/f6:data1/f7:data1/f8:data1/f1:data1/f2:data1/f3:data1/f4

[job_sdb2_sub5]
startdelay=0
rw=randread filename=data1/f6:data1/f7:data1/f8:data1/f1:data1/f2:data1/f3:data1/f4:data1/f5

[job_sdb2_sub6]
startdelay=0
rw=randread filename=data1/f7:data1/f8:data1/f1:data1/f2:data1/f3:data1/f4:data1/f5:data1/f6

[job_sdb2_sub7]
startdelay=0
rw=randread filename=data1/f8:data1/f1:data1/f2:data1/f3:data1/f4:data1/f5:data1/f6:data1/f7

[job_sdc1_sub0]
startdelay=0
rw=randread filename=data2/f1:data2/f2:data2/f3:data2/f4:data2/f5:data2/f6:data2/f7:data2/f8

[job_sdc1_sub1]
startdelay=0
rw=randread filename=data2/f2:data2/f3:data2/f4:data2/f5:data2/f6:data2/f7:data2/f8:data2/f1

[job_sdc1_sub2]
startdelay=0
rw=randread filename=data2/f3:data2/f4:data2/f5:data2/f6:data2/f7:data2/f8:data2/f1:data2/f2

[job_sdc1_sub3]
startdelay=0
rw=randread filename=data2/f4:data2/f5:data2/f6:data2/f7:data2/f8:data2/f1:data2/f2:data2/f3

[job_sdc1_sub4]
startdelay=0
rw=randread filename=data2/f5:data2/f6:data2/f7:data2/f8:data2/f1:data2/f2:data2/f3:data2/f4

[job_sdc1_sub5]
startdelay=0
rw=randread filename=data2/f6:data2/f7:data2/f8:data2/f1:data2/f2:data2/f3:data2/f4:data2/f5

[job_sdc1_sub6]
startdelay=0
rw=randread filename=data2/f7:data2/f8:data2/f1:data2/f2:data2/f3:data2/f4:data2/f5:data2/f6

[job_sdc1_sub7]
startdelay=0
rw=randread filename=data2/f8:data2/f1:data2/f2:data2/f3:data2/f4:data2/f5:data2/f6:data2/f7

[job_sdc2_sub0]
startdelay=0
rw=randread filename=data3/f1:data3/f2:data3/f3:data3/f4:data3/f5:data3/f6:data3/f7:data3/f8

[job_sdc2_sub1]
startdelay=0
rw=randread filename=data3/f2:data3/f3:data3/f4:data3/f5:data3/f6:data3/f7:data3/f8:data3/f1

[job_sdc2_sub2]
startdelay=0
rw=randread filename=data3/f3:data3/f4:data3/f5:data3/f6:data3/f7:data3/f8:data3/f1:data3/f2

[job_sdc2_sub3]
startdelay=0
rw=randread filename=data3/f4:data3/f5:data3/f6:data3/f7:data3/f8:data3/f1:data3/f2:data3/f3

[job_sdc2_sub4]
startdelay=0
rw=randread filename=data3/f5:data3/f6:data3/f7:data3/f8:data3/f1:data3/f2:data3/f3:data3/f4

[job_sdc2_sub5]
startdelay=0
rw=randread filename=data3/f6:data3/f7:data3/f8:data3/f1:data3/f2:data3/f3:data3/f4:data3/f5

[job_sdc2_sub6]
startdelay=0
rw=randread filename=data3/f7:data3/f8:data3/f1:data3/f2:data3/f3:data3/f4:data3/f5:data3/f6

[job_sdc2_sub7]
startdelay=0
rw=randread filename=data3/f8:data3/f1:data3/f2:data3/f3:data3/f4:data3/f5:data3/f6:data3/f7

[job_sdd1_sub0]
startdelay=0
rw=randread filename=data4/f1:data4/f2:data4/f3:data4/f4:data4/f5:data4/f6:data4/f7:data4/f8

[job_sdd1_sub1]
startdelay=0
rw=randread filename=data4/f2:data4/f3:data4/f4:data4/f5:data4/f6:data4/f7:data4/f8:data4/f1

[job_sdd1_sub2]
startdelay=0
rw=randread filename=data4/f3:data4/f4:data4/f5:data4/f6:data4/f7:data4/f8:data4/f1:data4/f2

[job_sdd1_sub3]
startdelay=0
rw=randread filename=data4/f4:data4/f5:data4/f6:data4/f7:data4/f8:data4/f1:data4/f2:data4/f3

[job_sdd1_sub4]
startdelay=0
rw=randread filename=data4/f5:data4/f6:data4/f7:data4/f8:data4/f1:data4/f2:data4/f3:data4/f4

[job_sdd1_sub5]
startdelay=0
rw=randread filename=data4/f6:data4/f7:data4/f8:data4/f1:data4/f2:data4/f3:data4/f4:data4/f5

[job_sdd1_sub6]
startdelay=0
rw=randread filename=data4/f7:data4/f8:data4/f1:data4/f2:data4/f3:data4/f4:data4/f5:data4/f6

[job_sdd1_sub7]
startdelay=0
rw=randread filename=data4/f8:data4/f1:data4/f2:data4/f3:data4/f4:data4/f5:data4/f6:data4/f7

[job_sdd2_sub0]
startdelay=0
rw=randread filename=data5/f1:data5/f2:data5/f3:data5/f4:data5/f5:data5/f6:data5/f7:data5/f8

[job_sdd2_sub1]
startdelay=0
rw=randread filename=data5/f2:data5/f3:data5/f4:data5/f5:data5/f6:data5/f7:data5/f8:data5/f1

[job_sdd2_sub2]
startdelay=0
rw=randread filename=data5/f3:data5/f4:data5/f5:data5/f6:data5/f7:data5/f8:data5/f1:data5/f2

[job_sdd2_sub3]
startdelay=0
rw=randread filename=data5/f4:data5/f5:data5/f6:data5/f7:data5/f8:data5/f1:data5/f2:data5/f3

[job_sdd2_sub4]
startdelay=0
rw=randread filename=data5/f5:data5/f6:data5/f7:data5/f8:data5/f1:data5/f2:data5/f3:data5/f4

[job_sdd2_sub5]
startdelay=0
rw=randread filename=data5/f6:data5/f7:data5/f8:data5/f1:data5/f2:data5/f3:data5/f4:data5/f5

[job_sdd2_sub6]
startdelay=0
rw=randread filename=data5/f7:data5/f8:data5/f1:data5/f2:data5/f3:data5/f4:data5/f5:data5/f6

[job_sdd2_sub7]
startdelay=0
rw=randread filename=data5/f8:data5/f1:data5/f2:data5/f3:data5/f4:data5/f5:data5/f6:data5/f7

[job_sde1_sub0]
startdelay=0
rw=randread filename=data6/f1:data6/f2:data6/f3:data6/f4:data6/f5:data6/f6:data6/f7:data6/f8

[job_sde1_sub1]
startdelay=0
rw=randread filename=data6/f2:data6/f3:data6/f4:data6/f5:data6/f6:data6/f7:data6/f8:data6/f1

[job_sde1_sub2]
startdelay=0
rw=randread filename=data6/f3:data6/f4:data6/f5:data6/f6:data6/f7:data6/f8:data6/f1:data6/f2

[job_sde1_sub3]
startdelay=0
rw=randread filename=data6/f4:data6/f5:data6/f6:data6/f7:data6/f8:data6/f1:data6/f2:data6/f3

[job_sde1_sub4]
startdelay=0
rw=randread filename=data6/f5:data6/f6:data6/f7:data6/f8:data6/f1:data6/f2:data6/f3:data6/f4

[job_sde1_sub5]
startdelay=0
rw=randread filename=data6/f6:data6/f7:data6/f8:data6/f1:data6/f2:data6/f3:data6/f4:data6/f5

[job_sde1_sub6]
startdelay=0
rw=randread filename=data6/f7:data6/f8:data6/f1:data6/f2:data6/f3:data6/f4:data6/f5:data6/f6

[job_sde1_sub7]
startdelay=0
rw=randread filename=data6/f8:data6/f1:data6/f2:data6/f3:data6/f4:data6/f5:data6/f6:data6/f7

[job_sde2_sub0]
startdelay=0
rw=randread filename=data7/f1:data7/f2:data7/f3:data7/f4:data7/f5:data7/f6:data7/f7:data7/f8

[job_sde2_sub1]
startdelay=0
rw=randread filename=data7/f2:data7/f3:data7/f4:data7/f5:data7/f6:data7/f7:data7/f8:data7/f1

[job_sde2_sub2]
startdelay=0
rw=randread filename=data7/f3:data7/f4:data7/f5:data7/f6:data7/f7:data7/f8:data7/f1:data7/f2

[job_sde2_sub3]
startdelay=0
rw=randread filename=data7/f4:data7/f5:data7/f6:data7/f7:data7/f8:data7/f1:data7/f2:data7/f3

[job_sde2_sub4]
startdelay=0
rw=randread filename=data7/f5:data7/f6:data7/f7:data7/f8:data7/f1:data7/f2:data7/f3:data7/f4

[job_sde2_sub5]
startdelay=0
rw=randread filename=data7/f6:data7/f7:data7/f8:data7/f1:data7/f2:data7/f3:data7/f4:data7/f5

[job_sde2_sub6]
startdelay=0
rw=randread filename=data7/f7:data7/f8:data7/f1:data7/f2:data7/f3:data7/f4:data7/f5:data7/f6

[job_sde2_sub7]
startdelay=0
rw=randread filename=data7/f8:data7/f1:data7/f2:data7/f3:data7/f4:data7/f5:data7/f6:data7/f7

[job_sdf1_sub0]
startdelay=0
rw=randread filename=data8/f1:data8/f2:data8/f3:data8/f4:data8/f5:data8/f6:data8/f7:data8/f8

[job_sdf1_sub1]
startdelay=0
rw=randread filename=data8/f2:data8/f3:data8/f4:data8/f5:data8/f6:data8/f7:data8/f8:data8/f1

[job_sdf1_sub2]
startdelay=0
rw=randread filename=data8/f3:data8/f4:data8/f5:data8/f6:data8/f7:data8/f8:data8/f1:data8/f2

[job_sdf1_sub3]
startdelay=0
rw=randread filename=data8/f4:data8/f5:data8/f6:data8/f7:data8/f8:data8/f1:data8/f2:data8/f3

[job_sdf1_sub4]
startdelay=0
rw=randread filename=data8/f5:data8/f6:data8/f7:data8/f8:data8/f1:data8/f2:data8/f3:data8/f4

[job_sdf1_sub5]
startdelay=0
rw=randread filename=data8/f6:data8/f7:data8/f8:data8/f1:data8/f2:data8/f3:data8/f4:data8/f5

[job_sdf1_sub6]
startdelay=0
rw=randread filename=data8/f7:data8/f8:data8/f1:data8/f2:data8/f3:data8/f4:data8/f5:data8/f6

[job_sdf1_sub7]
startdelay=0
rw=randread filename=data8/f8:data8/f1:data8/f2:data8/f3:data8/f4:data8/f5:data8/f6:data8/f7

[job_sdf2_sub0]
startdelay=0
rw=randread filename=data9/f1:data9/f2:data9/f3:data9/f4:data9/f5:data9/f6:data9/f7:data9/f8

[job_sdf2_sub1]
startdelay=0
rw=randread filename=data9/f2:data9/f3:data9/f4:data9/f5:data9/f6:data9/f7:data9/f8:data9/f1

[job_sdf2_sub2]
startdelay=0
rw=randread filename=data9/f3:data9/f4:data9/f5:data9/f6:data9/f7:data9/f8:data9/f1:data9/f2

[job_sdf2_sub3]
startdelay=0
rw=randread filename=data9/f4:data9/f5:data9/f6:data9/f7:data9/f8:data9/f1:data9/f2:data9/f3

[job_sdf2_sub4]
startdelay=0
rw=randread filename=data9/f5:data9/f6:data9/f7:data9/f8:data9/f1:data9/f2:data9/f3:data9/f4

[job_sdf2_sub5]
startdelay=0
rw=randread filename=data9/f6:data9/f7:data9/f8:data9/f1:data9/f2:data9/f3:data9/f4:data9/f5

[job_sdf2_sub6]
startdelay=0
rw=randread filename=data9/f7:data9/f8:data9/f1:data9/f2:data9/f3:data9/f4:data9/f5:data9/f6

[job_sdf2_sub7]
startdelay=0
rw=randread filename=data9/f8:data9/f1:data9/f2:data9/f3:data9/f4:data9/f5:data9/f6:data9/f7

[job_sdg1_sub0]
startdelay=0
rw=randread filename=data10/f1:data10/f2:data10/f3:data10/f4:data10/f5:data10/f6:data10/f7:data10/f8

[job_sdg1_sub1]
startdelay=0
rw=randread filename=data10/f2:data10/f3:data10/f4:data10/f5:data10/f6:data10/f7:data10/f8:data10/f1

[job_sdg1_sub2]
startdelay=0
rw=randread filename=data10/f3:data10/f4:data10/f5:data10/f6:data10/f7:data10/f8:data10/f1:data10/f2

[job_sdg1_sub3]
startdelay=0
rw=randread filename=data10/f4:data10/f5:data10/f6:data10/f7:data10/f8:data10/f1:data10/f2:data10/f3

[job_sdg1_sub4]
startdelay=0
rw=randread filename=data10/f5:data10/f6:data10/f7:data10/f8:data10/f1:data10/f2:data10/f3:data10/f4

[job_sdg1_sub5]
startdelay=0
rw=randread filename=data10/f6:data10/f7:data10/f8:data10/f1:data10/f2:data10/f3:data10/f4:data10/f5

[job_sdg1_sub6]
startdelay=0
rw=randread filename=data10/f7:data10/f8:data10/f1:data10/f2:data10/f3:data10/f4:data10/f5:data10/f6

[job_sdg1_sub7]
startdelay=0
rw=randread filename=data10/f8:data10/f1:data10/f2:data10/f3:data10/f4:data10/f5:data10/f6:data10/f7

[job_sdg2_sub0]
startdelay=0
rw=randread filename=data11/f1:data11/f2:data11/f3:data11/f4:data11/f5:data11/f6:data11/f7:data11/f8

[job_sdg2_sub1]
startdelay=0
rw=randread filename=data11/f2:data11/f3:data11/f4:data11/f5:data11/f6:data11/f7:data11/f8:data11/f1

[job_sdg2_sub2]
startdelay=0
rw=randread filename=data11/f3:data11/f4:data11/f5:data11/f6:data11/f7:data11/f8:data11/f1:data11/f2

[job_sdg2_sub3]
startdelay=0
rw=randread filename=data11/f4:data11/f5:data11/f6:data11/f7:data11/f8:data11/f1:data11/f2:data11/f3

[job_sdg2_sub4]
startdelay=0
rw=randread filename=data11/f5:data11/f6:data11/f7:data11/f8:data11/f1:data11/f2:data11/f3:data11/f4

[job_sdg2_sub5]
startdelay=0
rw=randread filename=data11/f6:data11/f7:data11/f8:data11/f1:data11/f2:data11/f3:data11/f4:data11/f5

[job_sdg2_sub6]
startdelay=0
rw=randread filename=data11/f7:data11/f8:data11/f1:data11/f2:data11/f3:data11/f4:data11/f5:data11/f6

[job_sdg2_sub7]
startdelay=0
rw=randread filename=data11/f8:data11/f1:data11/f2:data11/f3:data11/f4:data11/f5:data11/f6:data11/f7

[job_sdh1_sub0]
startdelay=0
rw=randread filename=data12/f1:data12/f2:data12/f3:data12/f4:data12/f5:data12/f6:data12/f7:data12/f8

[job_sdh1_sub1]
startdelay=0
rw=randread filename=data12/f2:data12/f3:data12/f4:data12/f5:data12/f6:data12/f7:data12/f8:data12/f1

[job_sdh1_sub2]
startdelay=0
rw=randread filename=data12/f3:data12/f4:data12/f5:data12/f6:data12/f7:data12/f8:data12/f1:data12/f2

[job_sdh1_sub3]
startdelay=0
rw=randread filename=data12/f4:data12/f5:data12/f6:data12/f7:data12/f8:data12/f1:data12/f2:data12/f3

[job_sdh1_sub4]
startdelay=0
rw=randread filename=data12/f5:data12/f6:data12/f7:data12/f8:data12/f1:data12/f2:data12/f3:data12/f4

[job_sdh1_sub5]
startdelay=0
rw=randread filename=data12/f6:data12/f7:data12/f8:data12/f1:data12/f2:data12/f3:data12/f4:data12/f5

[job_sdh1_sub6]
startdelay=0
rw=randread filename=data12/f7:data12/f8:data12/f1:data12/f2:data12/f3:data12/f4:data12/f5:data12/f6

[job_sdh1_sub7]
startdelay=0
rw=randread filename=data12/f8:data12/f1:data12/f2:data12/f3:data12/f4:data12/f5:data12/f6:data12/f7

[job_sdh2_sub0]
startdelay=0
rw=randread filename=data13/f1:data13/f2:data13/f3:data13/f4:data13/f5:data13/f6:data13/f7:data13/f8

[job_sdh2_sub1]
startdelay=0
rw=randread filename=data13/f2:data13/f3:data13/f4:data13/f5:data13/f6:data13/f7:data13/f8:data13/f1

[job_sdh2_sub2]
startdelay=0
rw=randread filename=data13/f3:data13/f4:data13/f5:data13/f6:data13/f7:data13/f8:data13/f1:data13/f2

[job_sdh2_sub3]
startdelay=0
rw=randread filename=data13/f4:data13/f5:data13/f6:data13/f7:data13/f8:data13/f1:data13/f2:data13/f3

[job_sdh2_sub4]
startdelay=0
rw=randread filename=data13/f5:data13/f6:data13/f7:data13/f8:data13/f1:data13/f2:data13/f3:data13/f4

[job_sdh2_sub5]
startdelay=0
rw=randread filename=data13/f6:data13/f7:data13/f8:data13/f1:data13/f2:data13/f3:data13/f4:data13/f5

[job_sdh2_sub6]
startdelay=0
rw=randread filename=data13/f7:data13/f8:data13/f1:data13/f2:data13/f3:data13/f4:data13/f5:data13/f6

[job_sdh2_sub7]
startdelay=0
rw=randread filename=data13/f8:data13/f1:data13/f2:data13/f3:data13/f4:data13/f5:data13/f6:data13/f7

[job_sdi1_sub0]
startdelay=0
rw=randread filename=data14/f1:data14/f2:data14/f3:data14/f4:data14/f5:data14/f6:data14/f7:data14/f8

[job_sdi1_sub1]
startdelay=0
rw=randread filename=data14/f2:data14/f3:data14/f4:data14/f5:data14/f6:data14/f7:data14/f8:data14/f1

[job_sdi1_sub2]
startdelay=0
rw=randread filename=data14/f3:data14/f4:data14/f5:data14/f6:data14/f7:data14/f8:data14/f1:data14/f2

[job_sdi1_sub3]
startdelay=0
rw=randread filename=data14/f4:data14/f5:data14/f6:data14/f7:data14/f8:data14/f1:data14/f2:data14/f3

[job_sdi1_sub4]
startdelay=0
rw=randread filename=data14/f5:data14/f6:data14/f7:data14/f8:data14/f1:data14/f2:data14/f3:data14/f4

[job_sdi1_sub5]
startdelay=0
rw=randread filename=data14/f6:data14/f7:data14/f8:data14/f1:data14/f2:data14/f3:data14/f4:data14/f5

[job_sdi1_sub6]
startdelay=0
rw=randread filename=data14/f7:data14/f8:data14/f1:data14/f2:data14/f3:data14/f4:data14/f5:data14/f6

[job_sdi1_sub7]
startdelay=0
rw=randread filename=data14/f8:data14/f1:data14/f2:data14/f3:data14/f4:data14/f5:data14/f6:data14/f7

[job_sdi2_sub0]
startdelay=0
rw=randread filename=data15/f1:data15/f2:data15/f3:data15/f4:data15/f5:data15/f6:data15/f7:data15/f8

[job_sdi2_sub1]
startdelay=0
rw=randread filename=data15/f2:data15/f3:data15/f4:data15/f5:data15/f6:data15/f7:data15/f8:data15/f1

[job_sdi2_sub2]
startdelay=0
rw=randread filename=data15/f3:data15/f4:data15/f5:data15/f6:data15/f7:data15/f8:data15/f1:data15/f2

[job_sdi2_sub3]
startdelay=0
rw=randread filename=data15/f4:data15/f5:data15/f6:data15/f7:data15/f8:data15/f1:data15/f2:data15/f3

[job_sdi2_sub4]
startdelay=0
rw=randread filename=data15/f5:data15/f6:data15/f7:data15/f8:data15/f1:data15/f2:data15/f3:data15/f4

[job_sdi2_sub5]
startdelay=0
rw=randread filename=data15/f6:data15/f7:data15/f8:data15/f1:data15/f2:data15/f3:data15/f4:data15/f5

[job_sdi2_sub6]
startdelay=0
rw=randread filename=data15/f7:data15/f8:data15/f1:data15/f2:data15/f3:data15/f4:data15/f5:data15/f6

[job_sdi2_sub7]
startdelay=0
rw=randread filename=data15/f8:data15/f1:data15/f2:data15/f3:data15/f4:data15/f5:data15/f6:data15/f7

[job_sdj1_sub0]
startdelay=0
rw=randread filename=data16/f1:data16/f2:data16/f3:data16/f4:data16/f5:data16/f6:data16/f7:data16/f8

[job_sdj1_sub1]
startdelay=0
rw=randread filename=data16/f2:data16/f3:data16/f4:data16/f5:data16/f6:data16/f7:data16/f8:data16/f1

[job_sdj1_sub2]
startdelay=0
rw=randread filename=data16/f3:data16/f4:data16/f5:data16/f6:data16/f7:data16/f8:data16/f1:data16/f2

[job_sdj1_sub3]
startdelay=0
rw=randread filename=data16/f4:data16/f5:data16/f6:data16/f7:data16/f8:data16/f1:data16/f2:data16/f3

[job_sdj1_sub4]
startdelay=0
rw=randread filename=data16/f5:data16/f6:data16/f7:data16/f8:data16/f1:data16/f2:data16/f3:data16/f4

[job_sdj1_sub5]
startdelay=0
rw=randread filename=data16/f6:data16/f7:data16/f8:data16/f1:data16/f2:data16/f3:data16/f4:data16/f5

[job_sdj1_sub6]
startdelay=0
rw=randread filename=data16/f7:data16/f8:data16/f1:data16/f2:data16/f3:data16/f4:data16/f5:data16/f6

[job_sdj1_sub7]
startdelay=0
rw=randread filename=data16/f8:data16/f1:data16/f2:data16/f3:data16/f4:data16/f5:data16/f6:data16/f7

[job_sdj2_sub0]
startdelay=0
rw=randread filename=data17/f1:data17/f2:data17/f3:data17/f4:data17/f5:data17/f6:data17/f7:data17/f8

[job_sdj2_sub1]
startdelay=0
rw=randread filename=data17/f2:data17/f3:data17/f4:data17/f5:data17/f6:data17/f7:data17/f8:data17/f1

[job_sdj2_sub2]
startdelay=0
rw=randread filename=data17/f3:data17/f4:data17/f5:data17/f6:data17/f7:data17/f8:data17/f1:data17/f2

[job_sdj2_sub3]
startdelay=0
rw=randread filename=data17/f4:data17/f5:data17/f6:data17/f7:data17/f8:data17/f1:data17/f2:data17/f3

[job_sdj2_sub4]
startdelay=0
rw=randread filename=data17/f5:data17/f6:data17/f7:data17/f8:data17/f1:data17/f2:data17/f3:data17/f4

[job_sdj2_sub5]
startdelay=0
rw=randread filename=data17/f6:data17/f7:data17/f8:data17/f1:data17/f2:data17/f3:data17/f4:data17/f5

[job_sdj2_sub6]
startdelay=0
rw=randread filename=data17/f7:data17/f8:data17/f1:data17/f2:data17/f3:data17/f4:data17/f5:data17/f6

[job_sdj2_sub7]
startdelay=0
rw=randread filename=data17/f8:data17/f1:data17/f2:data17/f3:data17/f4:data17/f5:data17/f6:data17/f7

[job_sdk1_sub0]
startdelay=0
rw=randread filename=data18/f1:data18/f2:data18/f3:data18/f4:data18/f5:data18/f6:data18/f7:data18/f8

[job_sdk1_sub1]
startdelay=0
rw=randread filename=data18/f2:data18/f3:data18/f4:data18/f5:data18/f6:data18/f7:data18/f8:data18/f1

[job_sdk1_sub2]
startdelay=0
rw=randread filename=data18/f3:data18/f4:data18/f5:data18/f6:data18/f7:data18/f8:data18/f1:data18/f2

[job_sdk1_sub3]
startdelay=0
rw=randread filename=data18/f4:data18/f5:data18/f6:data18/f7:data18/f8:data18/f1:data18/f2:data18/f3

[job_sdk1_sub4]
startdelay=0
rw=randread filename=data18/f5:data18/f6:data18/f7:data18/f8:data18/f1:data18/f2:data18/f3:data18/f4

[job_sdk1_sub5]
startdelay=0
rw=randread filename=data18/f6:data18/f7:data18/f8:data18/f1:data18/f2:data18/f3:data18/f4:data18/f5

[job_sdk1_sub6]
startdelay=0
rw=randread filename=data18/f7:data18/f8:data18/f1:data18/f2:data18/f3:data18/f4:data18/f5:data18/f6

[job_sdk1_sub7]
startdelay=0
rw=randread filename=data18/f8:data18/f1:data18/f2:data18/f3:data18/f4:data18/f5:data18/f6:data18/f7

[job_sdk2_sub0]
startdelay=0
rw=randread filename=data19/f1:data19/f2:data19/f3:data19/f4:data19/f5:data19/f6:data19/f7:data19/f8

[job_sdk2_sub1]
startdelay=0
rw=randread filename=data19/f2:data19/f3:data19/f4:data19/f5:data19/f6:data19/f7:data19/f8:data19/f1

[job_sdk2_sub2]
startdelay=0
rw=randread filename=data19/f3:data19/f4:data19/f5:data19/f6:data19/f7:data19/f8:data19/f1:data19/f2

[job_sdk2_sub3]
startdelay=0
rw=randread filename=data19/f4:data19/f5:data19/f6:data19/f7:data19/f8:data19/f1:data19/f2:data19/f3

[job_sdk2_sub4]
startdelay=0
rw=randread filename=data19/f5:data19/f6:data19/f7:data19/f8:data19/f1:data19/f2:data19/f3:data19/f4

[job_sdk2_sub5]
startdelay=0
rw=randread filename=data19/f6:data19/f7:data19/f8:data19/f1:data19/f2:data19/f3:data19/f4:data19/f5

[job_sdk2_sub6]
startdelay=0
rw=randread filename=data19/f7:data19/f8:data19/f1:data19/f2:data19/f3:data19/f4:data19/f5:data19/f6

[job_sdk2_sub7]
startdelay=0
rw=randread filename=data19/f8:data19/f1:data19/f2:data19/f3:data19/f4:data19/f5:data19/f6:data19/f7

[job_sdl1_sub0]
startdelay=0
rw=randread filename=data20/f1:data20/f2:data20/f3:data20/f4:data20/f5:data20/f6:data20/f7:data20/f8

[job_sdl1_sub1]
startdelay=0
rw=randread filename=data20/f2:data20/f3:data20/f4:data20/f5:data20/f6:data20/f7:data20/f8:data20/f1

[job_sdl1_sub2]
startdelay=0
rw=randread filename=data20/f3:data20/f4:data20/f5:data20/f6:data20/f7:data20/f8:data20/f1:data20/f2

[job_sdl1_sub3]
startdelay=0
rw=randread filename=data20/f4:data20/f5:data20/f6:data20/f7:data20/f8:data20/f1:data20/f2:data20/f3

[job_sdl1_sub4]
startdelay=0
rw=randread filename=data20/f5:data20/f6:data20/f7:data20/f8:data20/f1:data20/f2:data20/f3:data20/f4

[job_sdl1_sub5]
startdelay=0
rw=randread filename=data20/f6:data20/f7:data20/f8:data20/f1:data20/f2:data20/f3:data20/f4:data20/f5

[job_sdl1_sub6]
startdelay=0
rw=randread filename=data20/f7:data20/f8:data20/f1:data20/f2:data20/f3:data20/f4:data20/f5:data20/f6

[job_sdl1_sub7]
startdelay=0
rw=randread filename=data20/f8:data20/f1:data20/f2:data20/f3:data20/f4:data20/f5:data20/f6:data20/f7

[job_sdl2_sub0]
startdelay=0
rw=randread filename=data21/f1:data21/f2:data21/f3:data21/f4:data21/f5:data21/f6:data21/f7:data21/f8

[job_sdl2_sub1]
startdelay=0
rw=randread filename=data21/f2:data21/f3:data21/f4:data21/f5:data21/f6:data21/f7:data21/f8:data21/f1

[job_sdl2_sub2]
startdelay=0
rw=randread filename=data21/f3:data21/f4:data21/f5:data21/f6:data21/f7:data21/f8:data21/f1:data21/f2

[job_sdl2_sub3]
startdelay=0
rw=randread filename=data21/f4:data21/f5:data21/f6:data21/f7:data21/f8:data21/f1:data21/f2:data21/f3

[job_sdl2_sub4]
startdelay=0
rw=randread filename=data21/f5:data21/f6:data21/f7:data21/f8:data21/f1:data21/f2:data21/f3:data21/f4

[job_sdl2_sub5]
startdelay=0
rw=randread filename=data21/f6:data21/f7:data21/f8:data21/f1:data21/f2:data21/f3:data21/f4:data21/f5

[job_sdl2_sub6]
startdelay=0
rw=randread filename=data21/f7:data21/f8:data21/f1:data21/f2:data21/f3:data21/f4:data21/f5:data21/f6

[job_sdl2_sub7]
startdelay=0
rw=randread filename=data21/f8:data21/f1:data21/f2:data21/f3:data21/f4:data21/f5:data21/f6:data21/f7

[job_sdm1_sub0]
startdelay=0
rw=randread filename=data22/f1:data22/f2:data22/f3:data22/f4:data22/f5:data22/f6:data22/f7:data22/f8

[job_sdm1_sub1]
startdelay=0
rw=randread filename=data22/f2:data22/f3:data22/f4:data22/f5:data22/f6:data22/f7:data22/f8:data22/f1

[job_sdm1_sub2]
startdelay=0
rw=randread filename=data22/f3:data22/f4:data22/f5:data22/f6:data22/f7:data22/f8:data22/f1:data22/f2

[job_sdm1_sub3]
startdelay=0
rw=randread filename=data22/f4:data22/f5:data22/f6:data22/f7:data22/f8:data22/f1:data22/f2:data22/f3

[job_sdm1_sub4]
startdelay=0
rw=randread filename=data22/f5:data22/f6:data22/f7:data22/f8:data22/f1:data22/f2:data22/f3:data22/f4

[job_sdm1_sub5]
startdelay=0
rw=randread filename=data22/f6:data22/f7:data22/f8:data22/f1:data22/f2:data22/f3:data22/f4:data22/f5

[job_sdm1_sub6]
startdelay=0
rw=randread filename=data22/f7:data22/f8:data22/f1:data22/f2:data22/f3:data22/f4:data22/f5:data22/f6

[job_sdm1_sub7]
startdelay=0
rw=randread filename=data22/f8:data22/f1:data22/f2:data22/f3:data22/f4:data22/f5:data22/f6:data22/f7

[job_sdm2_sub0]
startdelay=0
rw=randread filename=data23/f1:data23/f2:data23/f3:data23/f4:data23/f5:data23/f6:data23/f7:data23/f8

[job_sdm2_sub1]
startdelay=0
rw=randread filename=data23/f2:data23/f3:data23/f4:data23/f5:data23/f6:data23/f7:data23/f8:data23/f1

[job_sdm2_sub2]
startdelay=0
rw=randread filename=data23/f3:data23/f4:data23/f5:data23/f6:data23/f7:data23/f8:data23/f1:data23/f2

[job_sdm2_sub3]
startdelay=0
rw=randread filename=data23/f4:data23/f5:data23/f6:data23/f7:data23/f8:data23/f1:data23/f2:data23/f3

[job_sdm2_sub4]
startdelay=0
rw=randread filename=data23/f5:data23/f6:data23/f7:data23/f8:data23/f1:data23/f2:data23/f3:data23/f4

[job_sdm2_sub5]
startdelay=0
rw=randread filename=data23/f6:data23/f7:data23/f8:data23/f1:data23/f2:data23/f3:data23/f4:data23/f5

[job_sdm2_sub6]
startdelay=0
rw=randread filename=data23/f7:data23/f8:data23/f1:data23/f2:data23/f3:data23/f4:data23/f5:data23/f6

[job_sdm2_sub7]
startdelay=0
rw=randread filename=data23/f8:data23/f1:data23/f2:data23/f3:data23/f4:data23/f5:data23/f6:data23/f7

#3 lore raw

Hi Yanmin,

Quoted 16 lines · Yanmin
Quoted 6 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines · Yanmin

Comparing with kernel 2.6.32, fio mmap randread 64k has more than 40% regression with 2.6.33-rc1.

Thanks for your timely reply. Some comments inlined below.

Quoted 1 line

Can you compare the performance also with 2.6.31?

We did. We run Linux kernel Performance Tracking project and run many benchmarks when a RC kernel is released.

The result of 2.6.31 is quite similar to the one of 2.6.32. But the one of 2.6.30 is about 8% better than the one of 2.6.31.

Quoted 6 lines
Quoted 4 lines

I think I understand what causes your problem.
2.6.32, with default settings, handled even random readers as sequential ones to provide fairness. This has benefits on single disks and JBODs, but causes harm on raids.

I didn't test RAID as that machine with hardware RAID HBA is crashed now. But if we turn on hardware RAID in HBA, mostly we use noop io scheduler.

I think you should start testing cfq with them, too. From 2.6.33, we have some big improvements in this area.

Quoted 58 lines
Quoted 11 lines

For 2.6.33, we changed the way in which this is handled, restoring the enable_idle = 0 for seeky queues as it was in 2.6.31:

(quoted diff) +0 -0
@@ -2218,13 +2352,10 @@ cfq_update_idle_window(struct cfq_data *cfqd,
 struct cfq_queue *cfqq,
        enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
 

       if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
-           (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
+           (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))) enable_idle = 0;

Quoted 8 lines

(compare with 2.6.31:
        if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || (cfqd->hw_tag && CIC_SEEKY(cic)))
                enable_idle = 0;
excluding the sample_valid check, it should be equivalent for you (I assume you have NCQ disks))
and we provide fairness for them by servicing all seeky queues together, and then idling before switching to other ones.

As for function cfq_update_idle_window, you is right. But since 2.6.32, CFQ merges many patches and the patches have impact on each other.

Quoted 36 lines

The mmap 64k randreader will have a large seek_mean, resulting in being marked seeky, but will send 16 * 4k sequential requests one after the other, so alternating between those seeky queues will cause harm.

I'm working on a new way to compute seekiness of queues, that should fix your issue, correctly identifying those queues as non-seeky (for me, a queue should be considered seeky only if it submits more than 1 seeky requests for 8 sequential ones).

Quoted 25 lines

The test scenario: 1 JBOD has 12 disks and every disk has 2 partitions. Create 8 1-GB files per partition and start 8 processes to do rand read on the 8 files per partitions. There are 8*24 processes totally. randread block size is 64K.

We found the regression on 2 machines. One machine has 8GB memory and the other has 6GB.

Bisect is very unstable. The related patches are many instead of just one.

1) commit 8e550632cccae34e265cb066691945515eaa7fb5
Author: Corrado Zoccolo <[email protected]>
Date:   Thu Nov 26 10:02:58 2009 +0100

   cfq-iosched: fix corner cases in idling logic

This patch introduces about less than 20% regression. I just reverted below section and this part regression disappear. It shows this regression is stable and not impacted by other patches.

(quoted diff) +0 -0
@@ -1253,9 +1254,9 @@ static void cfq_arm_slice_timer(struct cfq_data *cfqd)
                return;
 
Quoted 10 lines
Quoted 7 lines

Review static void cfq_arm_slice_timer(struct cfq_data *cfqd) · line 1260

Quoted 7 lines
(quoted diff) +2 -2
        /*
-        * still requests with the driver, don't idle
+        * still active requests from this queue, don't idle
         */
-       if (rq_in_driver(cfqd))
+       if (cfqq->dispatched)
                return;

Although 5 patches are related to the regression, above line is quite independent. Reverting above line could always improve the result for about 20%.

I've looked at your fio script, and it is quite complex, with lot of things going on.
Let's keep this for last.
I've created a smaller test, that already shows some regression:
[global]
direct=0
ioengine=mmap
size=8G
bs=64k
numjobs=1
loops=5
runtime=60
#group_reporting
invalidate=0
directory=/media/hd/cfq-tests

[job0]
startdelay=0
rw=randread
filename=testfile1

[job1]
startdelay=0
rw=randread
filename=testfile2

[job2]
startdelay=0
rw=randread
filename=testfile3

[job3]
startdelay=0
rw=randread
filename=testfile4

The attached patches, in particular 0005 (that apply on top of for-linus branch of Jen's tree
git://git.kernel.dk/linux-2.6-block.git) fix the regression on this simplified workload.

Quoted 4 lines
Quoted 3 lines
Quoted 2 lines

This shouldn't affect you if all queues are marked as idle.

Do you mean to use command ionice to mark it as idle class? I didn't try it.

No. I meant forcing enable_idle = 1, as you were almost doing with your patch, when cfq_latency was set.
With my above patch, this should not be needed any more, since the queues should be seen as sequential.

Quoted 7 lines
Quoted 7 lines

 Does just
your patch:

Quoted 4 lines

-           (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
-            && CFQQ_SEEKY(cfqq)))
+           (!cfqd->cfq_latency && !cfq_cfqq_deep(cfqq) &&
+               sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))

Quoted 5 lines
Quoted 1 line

fix most of the regression without touching arm_slice_timer?

No. If to fix the regression completely, I need apply above patch plus a debug patch. The debug patch is to just work around the 3 patches report by Shaohua's tiobench regression report. Without the debug patch, the regression isn't resolved.

Jens already merged one of Shaohua's patches, that may fix the problem with queue combining.

block/cfq-iosched.c    2009-12-30 17:12:28.000000000 +0800 cfq_set_prio_slice(struct cfq_data *cfqd · line 597

Quoted 11 lines

Below is the debug patch.
diff -Nraup linux-2.6.33_rc1/block/cfq-iosched.c linux-2.6.33_rc1_randread64k/block/cfq-iosched.c

block/cfq-iosched.c    2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c        2009-12-23 14:12:03.000000000 +0800 +3 -0
@@ -592,6 +592,9 @@ cfq_set_prio_slice(struct cfq_data *cfqd
        cfqq->slice_start = jiffies;
        cfqq->slice_end = jiffies + slice;
        cfqq->allocated_slice = slice;
+/*YMZHANG*/
+       cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
+

This is disabled, on a vanilla 2.6.33 kernel, by setting low_latency = 0

Quoted 14 lines
block/cfq-iosched.c    2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c        2009-12-23 14:12:03.000000000 +0800 +2 -1
        cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
  }
 
@@ -1836,7 +1839,8 @@ static void cfq_arm_slice_timer(struct c
        /*
         * still active requests from this queue, don't idle
         */
-       if (cfqq->dispatched)
+       //if (cfqq->dispatched)
+       if (rq_in_driver(cfqd))
                return;
 
        /*

block/cfq-iosched.c    2009-12-30 17:12:28.000000000 +0800 static void cfq_setup_merge(struct cfq_q · line 1950

Quoted 6 lines
block/cfq-iosched.c    2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c        2009-12-23 14:12:03.000000000 +0800 +3 -0
@@ -1941,6 +1945,9 @@ static void cfq_setup_merge(struct cfq_q
                new_cfqq = __cfqq;
        }
 
+       /* YMZHANG debug */
+       return;
+

This should be partially addressed by Shaohua's patch merged in Jens' tree. But note that your 8 processes, can randomly start doing I/O on the same file, so merging those queues is sometimes reasonable.
The patch to split them quickly was still not merged, though, so you will still see some regression due to this. In my simplified job file, I removed the randomness to make sure this cannot happen.

Quoted 8 lines
block/cfq-iosched.c    2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c        2009-12-23 14:12:03.000000000 +0800 +0 -0
        process_refs = cfqq_process_refs(cfqq);
        /*
         * If the process for the cfqq has gone away, there is no
Quoted 5 lines

I guess

Quoted 1 line

5db5d64277bf390056b1a87d0bb288c8b8553f96.

Quoted 8 lines
Quoted 2 lines

will still introduce a 10% regression, but this is needed to improve latency, and you can just disable low_latency to avoid it.

You are right. I did a quick testing. If my patch + revert 2 patches and keep 5db5d64, the regression is about 20%.

But low_latency=0 doesn't work like what we imagined. If patch + revert 2 patches and keep 5db5d64 while set low_latency=0, the regression is still there. One reason is my patch doesn't work when low_latency=0.

Right. You can try with my patch, instead, that doesn't depend on low_latency, and set it to 0 to remove this performance degradation. My results:
2.6.32.2:
READ: io=146688KB, aggrb=2442KB/s, minb=602KB/s, maxb=639KB/s,
mint=60019msec, maxt=60067msec

2.6.33 - jens:
READ: io=128512KB, aggrb=2140KB/s, minb=526KB/s, maxb=569KB/s,
mint=60004msec, maxt=60032msec

2.6.33 - jens + my patches :
READ: io=143232KB, aggrb=2384KB/s, minb=595KB/s, maxb=624KB/s,
mint=60003msec, maxt=60072msec

2.6.33 - jens + my patches + low_lat = 0:
READ: io=145216KB, aggrb=2416KB/s, minb=596KB/s, maxb=632KB/s,
mint=60027msec, maxt=60087msec

Quoted 9 lines
Quoted 3 lines

Thanks,
Corrado

I attach the fio job file for your reference.

I got a cold and will continue to work on it next week.

Yanmin

Thanks,
Corrado

From f7bf4db76818d6a0ce83c179cb6ae2a305af3082 Mon Sep 17 00:00:00 2001
From: Corrado Zoccolo <[email protected]>
Date: Wed, 30 Dec 2009 11:58:34 +0100
Subject: [PATCH 3/5] cfq-iosched: non-rot devices do not need read queue merging

Non rotational devices' performances are not affected by distance of read requests, so there is no point in having overhead to merge such queues.
This doesn't apply to writes, so this patch changes the queued[] field, to be indexed by READ/WRITE instead of SYNC/ASYNC, and only compute proximity for queues with WRITE requests.

---
 block/cfq-iosched.c |   20 +++++++++++---------
 1 files changed, 11 insertions(+), 9 deletions(-)
block/cfq-iosched.c +11 -9
index 918c7fd..7da9391 100644
@@ -108,9 +108,9 @@ struct cfq_queue {
 	struct rb_root sort_list;
 	/* if fifo isn't expired, next request to serve */
 	struct request *next_rq;
-	/* requests queued in sort_list */
+	/* requests queued in sort_list, indexed by READ/WRITE */
 	int queued[2];
-	/* currently allocated requests */
+	/* currently allocated requests, indexed by READ/WRITE */
 	int allocated[2];
 	/* fifo list of requests in sort_list */
 	struct list_head fifo;
@@ -1268,7 +1268,8 @@ static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 		return;
 	if (!cfqq->next_rq)
 		return;
-
+	if (blk_queue_nonrot(cfqd->queue) && !cfqq->queued[WRITE])
+		return;
 	cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
 	__cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
 				      blk_rq_pos(cfqq->next_rq), &parent, &p);
@@ -1337,10 +1338,10 @@ static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 static void cfq_del_rq_rb(struct request *rq)
 {
 	struct cfq_queue *cfqq = RQ_CFQQ(rq);
-	const int sync = rq_is_sync(rq);
+	const int rw = rq_data_dir(rq);
 
-	BUG_ON(!cfqq->queued[sync]);
-	cfqq->queued[sync]--;
+	BUG_ON(!cfqq->queued[rw]);
+	cfqq->queued[rw]--;
 
 	elv_rb_del(&cfqq->sort_list, rq);
 
@@ -1363,7 +1364,7 @@ static void cfq_add_rq_rb(struct request *rq)
 	struct cfq_data *cfqd = cfqq->cfqd;
 	struct request *__alias, *prev;
 
-	cfqq->queued[rq_is_sync(rq)]++;
+	cfqq->queued[rq_data_dir(rq)]++;
 
 	/*
 	 * looks a little odd, but the first insert might return an alias.
@@ -1393,7 +1394,7 @@ static void cfq_add_rq_rb(struct request *rq)
 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
 {
 	elv_rb_del(&cfqq->sort_list, rq);
-	cfqq->queued[rq_is_sync(rq)]--;
+	cfqq->queued[rq_data_dir(rq)]--;
 	cfq_add_rq_rb(rq);
 }
 
@@ -1689,7 +1690,8 @@ static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
 	struct cfq_queue *__cfqq;
 	sector_t sector = cfqd->last_position;
 
-	if (RB_EMPTY_ROOT(root))
+	if (RB_EMPTY_ROOT(root) ||
+	    (blk_queue_nonrot(cfqd->queue) && !cur_cfqq->queued[WRITE]))
 		return NULL;
 
 	/*
Signature · 1 line
1.6.4.4

From bd47454a4381f584e79b9bb57eb0329e4b385ee5 Mon Sep 17 00:00:00 2001
From: Corrado Zoccolo <[email protected]>
Date: Wed, 30 Dec 2009 22:49:42 +0100
Subject: [PATCH 4/5] cfq-iosched: requests "in flight" vs "in driver" clarification

Counters for requests "in flight" and "in driver" are used asymmetrically in cfq_may_dispatch, and have slightly different meaning.
We split the rq_in_flight counter (was sync_flight) to count both sync and async requests, in order to use this one, which is more accurate in some corner cases.
The rq_in_driver counter is coalesced, since individual sync/async counts are not used any more.

---
 block/cfq-iosched.c |   44 ++++++++++++++++++--------------------------
 1 files changed, 18 insertions(+), 26 deletions(-)
block/cfq-iosched.c +18 -26

120 lines not on this page · raw message

Signature · 1 line
1.6.4.4

From c6eb136205c0b6ebe2e9732de249ddefba26d41d Mon Sep 17 00:00:00 2001
From: Corrado Zoccolo <[email protected]>
Date: Thu, 31 Dec 2009 13:41:59 +0100
Subject: [PATCH 5/5] cfq-iosched: rework seeky detection

---
 block/cfq-iosched.c |   54 +++++++++++++-------------------------------------
 1 files changed, 14 insertions(+), 40 deletions(-)
block/cfq-iosched.c +14 -40

83 lines not on this page · raw message

Signature · 1 line
1.6.4.4
#4 lore raw
Quoted 19 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 16 lines · Yanmin
Quoted 6 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines · Yanmin

Comparing with kernel 2.6.32, fio mmap randread 64k has more than 40% regression with 2.6.33-rc1.

Thanks for your timely reply. Some comments inlined below.

Quoted 1 line

Can you compare the performance also with 2.6.31?

We did. We run Linux kernel Performance Tracking project and run many benchmarks when a RC kernel is released.

The result of 2.6.31 is quite similar to the one of 2.6.32. But the one of 2.6.30 is about 8% better than the one of 2.6.31.

Quoted 8 lines
Quoted 6 lines
Quoted 4 lines

I think I understand what causes your problem.
2.6.32, with default settings, handled even random readers as sequential ones to provide fairness. This has benefits on single disks and JBODs, but causes harm on raids.

I didn't test RAID as that machine with hardware RAID HBA is crashed now. But if we turn on hardware RAID in HBA, mostly we use noop io scheduler.

I think you should start testing cfq with them, too. From 2.6.33, we have some big improvements in this area.

Great! I once compared cfq and noop against non-raid and raid0. One interesting finding about sequential read testing is when there are fewer processes to read files on the raid0 JBOD, noop on raid0 is pretty good, but when there are lots of processes to do so on a non-raid JBOD, cfq is pretty better. I planed to investigate it, but too busy in other issues.

Quoted 68 lines
Quoted 58 lines
Quoted 12 lines

For 2.6.33, we changed the way in which this is handled, restoring the enable_idle = 0 for seeky queues as it was in 2.6.31:

(quoted diff) +1 -1
@@ -2218,13 +2352,10 @@ cfq_update_idle_window(struct cfq_data *cfqd,
 struct cfq_queue *cfqq,
       enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
 
       if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
-           (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
+           (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))
               enable_idle = 0;

(compare with 2.6.31:

Quoted 7 lines
        if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
            (cfqd->hw_tag && CIC_SEEKY(cic)))
                enable_idle = 0;

excluding the sample_valid check, it should be equivalent for you (I assume you have NCQ disks))
and we provide fairness for them by servicing all seeky queues together, and then idling before switching to other ones.

As for function cfq_update_idle_window, you is right. But since 2.6.32, CFQ merges many patches and the patches have impact on each other.

Quoted 36 lines

The mmap 64k randreader will have a large seek_mean, resulting in being marked seeky, but will send 16 * 4k sequential requests one after the other, so alternating between those seeky queues will cause harm.

I'm working on a new way to compute seekiness of queues, that should fix your issue, correctly identifying those queues as non-seeky (for me, a queue should be considered seeky only if it submits more than 1 seeky requests for 8 sequential ones).

Quoted 25 lines

The test scenario: 1 JBOD has 12 disks and every disk has 2 partitions. Create 8 1-GB files per partition and start 8 processes to do rand read on the 8 files per partitions. There are 8*24 processes totally. randread block size is 64K.

We found the regression on 2 machines. One machine has 8GB memory and the other has 6GB.

Bisect is very unstable. The related patches are many instead of just one.

1) commit 8e550632cccae34e265cb066691945515eaa7fb5

Author: Corrado Zoccolo <[email protected]>
Date:   Thu Nov 26 10:02:58 2009 +0100

cfq-iosched: fix corner cases in idling logic

This patch introduces about less than 20% regression. I just reverted below section and this part regression disappear. It shows this regression is stable and not impacted by other patches.

(quoted diff) +0 -0
@@ -1253,9 +1254,9 @@ static void cfq_arm_slice_timer(struct cfq_data *cfqd)
               return;
 
Quoted 10 lines
Quoted 7 lines

Review static void cfq_arm_slice_timer(struct cfq_data *cfqd) · line 1260

Quoted 7 lines
(quoted diff) +2 -2
       /*
-        * still requests with the driver, don't idle
+        * still active requests from this queue, don't idle
        */
-       if (rq_in_driver(cfqd))
+       if (cfqq->dispatched)
               return;

Although 5 patches are related to the regression, above line is quite independent. Reverting above line could always improve the result for about 20%.

Quoted 1 line

I've looked at your fio script, and it is quite complex,

As we have about 40 fio sub cases, we have a script to create fio job file from a specific parameter list. So there are some superfluous parameters.

Another point is we need stable result.

Quoted 3 lines

with lot of
things going on.
Let's keep this for last.

Ok. But the change like what you do mostly reduces regresion.

Quoted 33 lines

I've created a smaller test, that already shows some regression:
[global]
direct=0
ioengine=mmap
size=8G
bs=64k
numjobs=1
loops=5
runtime=60
#group_reporting
invalidate=0
directory=/media/hd/cfq-tests

[job0]
startdelay=0
rw=randread
filename=testfile1

[job1]
startdelay=0
rw=randread
filename=testfile2

[job2]
startdelay=0
rw=randread
filename=testfile3

[job3]
startdelay=0
rw=randread
filename=testfile4

Quoted 4 lines

The attached patches, in particular 0005 (that apply on top of for-linus branch of Jen's tree
git://git.kernel.dk/linux-2.6-block.git) fix the regression on this simplified workload.

I didn't download the tree. I tested the 3 attached patches against 2.6.33-rc1. The result isn't resolved.

Quoted 17 lines
Quoted 5 lines
Quoted 4 lines
Quoted 3 lines

This shouldn't affect you if all queues are marked as idle.

Do you mean to use command ionice to mark it as idle class? I didn't try it.

No. I meant forcing enable_idle = 1, as you were almost doing with your patch, when cfq_latency was set.
With my above patch, this should not be needed any more, since the queues should be seen as sequential.

Quoted 7 lines
Quoted 7 lines

Does just
your patch:

Quoted 4 lines
block/cfq-iosched.c +2 -2
-           (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
-            && CFQQ_SEEKY(cfqq)))
+           (!cfqd->cfq_latency && !cfq_cfqq_deep(cfqq) &&
+               sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))
Quoted 8 lines
Quoted 5 lines
Quoted 1 line

fix most of the regression without touching arm_slice_timer?

No. If to fix the regression completely, I need apply above patch plus a debug patch. The debug patch is to just work around the 3 patches report by Shaohua's tiobench regression report. Without the debug patch, the regression isn't resolved.

Jens already merged one of Shaohua's patches, that may fix the problem with queue combining.

I did another testing. Apply my debug patch+ the low_latency patch, but use Shaohua's 2 patches (improve merge and split), the regression disappears.

Quoted 30 lines

block/cfq-iosched.c 2009-12-30 17:12:28.000000000 +0800 cfq_set_prio_slice(struct cfq_data *cfqd · line 597

Quoted 12 lines

Below is the debug patch.
diff -Nraup linux-2.6.33_rc1/block/cfq-iosched.c linux-2.6.33_rc1_randread64k/block/cfq-iosched.c

block/cfq-iosched.c 2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c 2009-12-23 14:12:03.000000000 +0800 +3 -0
@@ -592,6 +592,9 @@ cfq_set_prio_slice(struct cfq_data *cfqd
       cfqq->slice_start = jiffies;
       cfqq->slice_end = jiffies + slice;
       cfqq->allocated_slice = slice;
+/*YMZHANG*/
+       cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
+

This is disabled, on a vanilla 2.6.33 kernel, by setting low_latency = 0

Quoted 17 lines
block/cfq-iosched.c 2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c 2009-12-23 14:12:03.000000000 +0800 +2 -1
       cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
 }
 
@@ -1836,7 +1839,8 @@ static void cfq_arm_slice_timer(struct c
       /*
        * still active requests from this queue, don't idle
        */
-       if (cfqq->dispatched)
+       //if (cfqq->dispatched)
+       if (rq_in_driver(cfqd))
               return;
 
       /*
@@ -1941,6 +1945,9 @@ static void cfq_setup_merge(struct cfq_q
               new_cfqq = __cfqq;
       }
 
Quoted 6 lines

block/cfq-iosched.c 2009-12-30 17:12:28.000000000 +0800 static void cfq_setup_merge(struct cfq_q · line 1950

Quoted 3 lines
block/cfq-iosched.c 2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c 2009-12-23 14:12:03.000000000 +0800 +3 -0
+       /* YMZHANG debug */
+       return;
+

This should be partially addressed by Shaohua's patch merged in Jens' tree. But note that your 8 processes, can randomly start doing I/O on the same file, so merging those queues is sometimes reasonable.

Another reason is I start 8 processes per partition and every disk has 2 partitions, so there are 16 processes per disk. With another JBOD, I use one partition per disk, and the regression is only 8%.

Quoted 1 line

From this point, can CFQ do not merge request queues which access different partitions?

As you know, it's unusual that a process accesses files across partitions. io scheduler is at low layer which doesn't know partition.

Quoted 52 lines

The patch to split them quickly was still not merged, though, so you will still see some regression due to this. In my simplified job file, I removed the randomness to make sure this cannot happen.

Quoted 8 lines
block/cfq-iosched.c 2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c 2009-12-23 14:12:03.000000000 +0800 +0 -0
       process_refs = cfqq_process_refs(cfqq);
       /*
        * If the process for the cfqq has gone away, there is no
Quoted 5 lines

I guess

Quoted 1 line

5db5d64277bf390056b1a87d0bb288c8b8553f96.

Quoted 8 lines
Quoted 2 lines

will still introduce a 10% regression, but this is needed to improve latency, and you can just disable low_latency to avoid it.

You are right. I did a quick testing. If my patch + revert 2 patches and keep 5db5d64, the regression is about 20%.

But low_latency=0 doesn't work like what we imagined. If patch + revert 2 patches and keep 5db5d64 while set low_latency=0, the regression is still there. One reason is my patch doesn't work when low_latency=0.

Right. You can try with my patch, instead, that doesn't depend on low_latency, and set it to 0 to remove this performance degradation. My results:
2.6.32.2:

   READ: io=146688KB, aggrb=2442KB/s, minb=602KB/s, maxb=639KB/s,

mint=60019msec, maxt=60067msec

2.6.33 - jens:

   READ: io=128512KB, aggrb=2140KB/s, minb=526KB/s, maxb=569KB/s,

mint=60004msec, maxt=60032msec

2.6.33 - jens + my patches :

   READ: io=143232KB, aggrb=2384KB/s, minb=595KB/s, maxb=624KB/s,

mint=60003msec, maxt=60072msec

2.6.33 - jens + my patches + low_lat = 0:

   READ: io=145216KB, aggrb=2416KB/s, minb=596KB/s, maxb=632KB/s,

mint=60027msec, maxt=60087msec

Quoted 9 lines
Quoted 3 lines

Thanks,
Corrado

I attach the fio job file for your reference.

I got a cold and will continue to work on it next week.

Yanmin

Thanks,
Corrado

#5 lore raw

Hi

Quoted 101 lines · Yanmin
Quoted 19 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 16 lines · Yanmin
Quoted 6 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines · Yanmin

Comparing with kernel 2.6.32, fio mmap randread 64k has more than 40% regression with 2.6.33-rc1.

Thanks for your timely reply. Some comments inlined below.

Quoted 1 line

Can you compare the performance also with 2.6.31?

We did. We run Linux kernel Performance Tracking project and run many benchmarks when a RC kernel is released.

The result of 2.6.31 is quite similar to the one of 2.6.32. But the one of 2.6.30 is about 8% better than the one of 2.6.31.

Quoted 8 lines
Quoted 6 lines
Quoted 4 lines

I think I understand what causes your problem.
2.6.32, with default settings, handled even random readers as sequential ones to provide fairness. This has benefits on single disks and JBODs, but causes harm on raids.

I didn't test RAID as that machine with hardware RAID HBA is crashed now. But if we turn on hardware RAID in HBA, mostly we use noop io scheduler.

I think you should start testing cfq with them, too. From 2.6.33, we have some big improvements in this area.

Great! I once compared cfq and noop against non-raid and raid0. One interesting finding about sequential read testing is when there are fewer processes to read files on the raid0 JBOD, noop on raid0 is pretty good, but when there are lots of processes to do so on a non-raid JBOD, cfq is pretty better. I planed to investigate it, but too busy in other issues.

Quoted 68 lines
Quoted 68 lines
Quoted 11 lines

For 2.6.33, we changed the way in which this is handled, restoring the enable_idle = 0 for seeky queues as it was in 2.6.31:

(quoted diff) +0 -0
@@ -2218,13 +2352,10 @@ cfq_update_idle_window(struct cfq_data *cfqd,
 struct cfq_queue *cfqq,
        enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
 

       if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
-           (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
+           (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))) enable_idle = 0;

Quoted 8 lines

(compare with 2.6.31:
        if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || (cfqd->hw_tag && CIC_SEEKY(cic)))
                enable_idle = 0;
excluding the sample_valid check, it should be equivalent for you (I assume you have NCQ disks))
and we provide fairness for them by servicing all seeky queues together, and then idling before switching to other ones.

As for function cfq_update_idle_window, you is right. But since 2.6.32, CFQ merges many patches and the patches have impact on each other.

Quoted 36 lines

The mmap 64k randreader will have a large seek_mean, resulting in being marked seeky, but will send 16 * 4k sequential requests one after the other, so alternating between those seeky queues will cause harm.

I'm working on a new way to compute seekiness of queues, that should fix your issue, correctly identifying those queues as non-seeky (for me, a queue should be considered seeky only if it submits more than 1 seeky requests for 8 sequential ones).

Quoted 25 lines

The test scenario: 1 JBOD has 12 disks and every disk has 2 partitions. Create 8 1-GB files per partition and start 8 processes to do rand read on the 8 files per partitions. There are 8*24 processes totally. randread block size is 64K.

We found the regression on 2 machines. One machine has 8GB memory and the other has 6GB.

Bisect is very unstable. The related patches are many instead of just one.

1) commit 8e550632cccae34e265cb066691945515eaa7fb5
Author: Corrado Zoccolo <[email protected]>
Date:   Thu Nov 26 10:02:58 2009 +0100

   cfq-iosched: fix corner cases in idling logic

This patch introduces about less than 20% regression. I just reverted below section and this part regression disappear. It shows this regression is stable and not impacted by other patches.

(quoted diff) +0 -0
@@ -1253,9 +1254,9 @@ static void cfq_arm_slice_timer(struct cfq_data *cfqd)
                return;
 
Quoted 7 lines

Review static void cfq_arm_slice_timer(struct cfq_data *cfqd) · line 1260

Quoted 7 lines
(quoted diff) +2 -2
        /*
-        * still requests with the driver, don't idle
+        * still active requests from this queue, don't idle
         */
-       if (rq_in_driver(cfqd))
+       if (cfqq->dispatched)
                return;

Although 5 patches are related to the regression, above line is quite independent. Reverting above line could always improve the result for about 20%.

Quoted 4 lines
Quoted 1 line

I've looked at your fio script, and it is quite complex,

As we have about 40 fio sub cases, we have a script to create fio job file from a specific parameter list. So there are some superfluous parameters.

My point is that there are so many things going on, that is more difficult to analyse the issues.
I prefer looking at one problem at a time, so (initially) removing the possibility of queue merging, that Shaohua already investigated, can help in spotting the still not-well-understood problem. Could you generate the same script, but with each process accessing only one of the files, instead of chosing it at random?

Quoted 40 lines

Another point is we need stable result.

Quoted 3 lines

with lot of
things going on.
Let's keep this for last.

Ok. But the change like what you do mostly reduces regresion.

Quoted 33 lines

I've created a smaller test, that already shows some regression:
[global]
direct=0
ioengine=mmap
size=8G
bs=64k
numjobs=1
loops=5
runtime=60
#group_reporting
invalidate=0
directory=/media/hd/cfq-tests

[job0]
startdelay=0
rw=randread
filename=testfile1

[job1]
startdelay=0
rw=randread
filename=testfile2

[job2]
startdelay=0
rw=randread
filename=testfile3

[job3]
startdelay=0
rw=randread
filename=testfile4

Quoted 6 lines
Quoted 4 lines

The attached patches, in particular 0005 (that apply on top of for-linus branch of Jen's tree
git://git.kernel.dk/linux-2.6-block.git) fix the regression on this simplified workload.

I didn't download the tree. I tested the 3 attached patches against 2.6.33-rc1. The result isn't resolved.

Can you quantify if there is an improvement, though?
Please, also include Shahoua's patches.
I'd like to see the comparison between (always with low_latency set to 0):
plain 2.6.33
plain 2.6.33 + shahoua's
plain 2.6.33 + shahoua's + my patch
plain 2.6.33 + shahoua's + my patch + rq_in_driver vs dispatched patch.

Quoted 64 lines
Quoted 17 lines
Quoted 5 lines
Quoted 4 lines
Quoted 3 lines

This shouldn't affect you if all queues are marked as idle.

Do you mean to use command ionice to mark it as idle class? I didn't try it.

No. I meant forcing enable_idle = 1, as you were almost doing with your patch, when cfq_latency was set.
With my above patch, this should not be needed any more, since the queues should be seen as sequential.

Quoted 7 lines
Quoted 7 lines

 Does just
your patch:

Quoted 4 lines

-           (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
-            && CFQQ_SEEKY(cfqq)))
+           (!cfqd->cfq_latency && !cfq_cfqq_deep(cfqq) &&
+               sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))

Quoted 8 lines
Quoted 5 lines
Quoted 1 line

fix most of the regression without touching arm_slice_timer?

No. If to fix the regression completely, I need apply above patch plus a debug patch. The debug patch is to just work around the 3 patches report by Shaohua's tiobench regression report. Without the debug patch, the regression isn't resolved.

Jens already merged one of Shaohua's patches, that may fix the problem with queue combining.

I did another testing. Apply my debug patch+ the low_latency patch, but use Shaohua's 2 patches (improve merge and split), the regression disappears.

Quoted 30 lines

block/cfq-iosched.c    2009-12-30 17:12:28.000000000 +0800 cfq_set_prio_slice(struct cfq_data *cfqd · line 597

Quoted 12 lines

Below is the debug patch.
diff -Nraup linux-2.6.33_rc1/block/cfq-iosched.c linux-2.6.33_rc1_randread64k/block/cfq-iosched.c

block/cfq-iosched.c    2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c        2009-12-23 14:12:03.000000000 +0800 +3 -0
@@ -592,6 +592,9 @@ cfq_set_prio_slice(struct cfq_data *cfqd
        cfqq->slice_start = jiffies;
        cfqq->slice_end = jiffies + slice;
        cfqq->allocated_slice = slice;
+/*YMZHANG*/
+       cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
+

This is disabled, on a vanilla 2.6.33 kernel, by setting low_latency = 0

Quoted 17 lines
block/cfq-iosched.c    2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c        2009-12-23 14:12:03.000000000 +0800 +2 -1
        cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
  }
 
@@ -1836,7 +1839,8 @@ static void cfq_arm_slice_timer(struct c
        /*
         * still active requests from this queue, don't idle
         */
-       if (cfqq->dispatched)
+       //if (cfqq->dispatched)
+       if (rq_in_driver(cfqd))
                return;
 
        /*
@@ -1941,6 +1945,9 @@ static void cfq_setup_merge(struct cfq_q
                new_cfqq = __cfqq;
        }
 
Quoted 6 lines

block/cfq-iosched.c    2009-12-30 17:12:28.000000000 +0800 static void cfq_setup_merge(struct cfq_q · line 1950

Quoted 3 lines
block/cfq-iosched.c    2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c        2009-12-23 14:12:03.000000000 +0800 +3 -0
+       /* YMZHANG debug */
+       return;
+

This should be partially addressed by Shaohua's patch merged in Jens' tree. But note that your 8 processes, can randomly start doing I/O on the same file, so merging those queues is sometimes reasonable.

Quoted 3 lines

Another reason is I start 8 processes per partition and every disk has 2 partitions, so there are 16 processes per disk. With another JBOD, I use one partition per disk, and the regression is only 8%.

With half of the processes, time slices are higher, and the disk cache can do a better job when servicing interleaved sequential requests.

Quoted 2 lines
Quoted 2 lines

From this point, can CFQ do not merge request queues which access different partitions?

(puzzled: I didn't write this, and can't find a message in the thread with this question.)

Quoted 2 lines

As you know, it's unusual that a process accesses files across partitions. io scheduler is at low layer which doesn't know partition.

CFQ bases decision on distance between requests, and requests going to different partitions will have much higher distance. So the associated queues will be more likely marked as seeky.

Quoted 57 lines
Quoted 54 lines

The patch to split them quickly was still not merged, though, so you will still see some regression due to this. In my simplified job file, I removed the randomness to make sure this cannot happen.

Quoted 8 lines
block/cfq-iosched.c    2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c        2009-12-23 14:12:03.000000000 +0800 +0 -0
        process_refs = cfqq_process_refs(cfqq);
        /*
         * If the process for the cfqq has gone away, there is no
Quoted 5 lines

I guess

Quoted 1 line

5db5d64277bf390056b1a87d0bb288c8b8553f96.

Quoted 8 lines
Quoted 2 lines

will still introduce a 10% regression, but this is needed to improve latency, and you can just disable low_latency to avoid it.

You are right. I did a quick testing. If my patch + revert 2 patches and keep 5db5d64, the regression is about 20%.

But low_latency=0 doesn't work like what we imagined. If patch + revert 2 patches and keep 5db5d64 while set low_latency=0, the regression is still there. One reason is my patch doesn't work when low_latency=0.

Right. You can try with my patch, instead, that doesn't depend on low_latency, and set it to 0 to remove this performance degradation. My results:
2.6.32.2:
   READ: io=146688KB, aggrb=2442KB/s, minb=602KB/s, maxb=639KB/s, mint=60019msec, maxt=60067msec

2.6.33 - jens:
   READ: io=128512KB, aggrb=2140KB/s, minb=526KB/s, maxb=569KB/s, mint=60004msec, maxt=60032msec

2.6.33 - jens + my patches :
   READ: io=143232KB, aggrb=2384KB/s, minb=595KB/s, maxb=624KB/s, mint=60003msec, maxt=60072msec

2.6.33 - jens + my patches + low_lat = 0:
   READ: io=145216KB, aggrb=2416KB/s, minb=596KB/s, maxb=632KB/s, mint=60027msec, maxt=60087msec

Quoted 9 lines
Quoted 3 lines

Thanks,
Corrado

I attach the fio job file for your reference.

I got a cold and will continue to work on it next week.

Yanmin

Thanks,
Corrado

#6 lore raw
Quoted 105 lines · Corrado Zoccolo

Hi

Quoted 102 lines · Yanmin
Quoted 19 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 16 lines · Yanmin
Quoted 6 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines · Yanmin

Comparing with kernel 2.6.32, fio mmap randread 64k has more than 40% regression with 2.6.33-rc1.

Thanks for your timely reply. Some comments inlined below.

Quoted 1 line

Can you compare the performance also with 2.6.31?

We did. We run Linux kernel Performance Tracking project and run many benchmarks when a RC kernel is released.

The result of 2.6.31 is quite similar to the one of 2.6.32. But the one of 2.6.30 is about 8% better than the one of 2.6.31.

Quoted 8 lines
Quoted 6 lines
Quoted 4 lines

I think I understand what causes your problem.
2.6.32, with default settings, handled even random readers as sequential ones to provide fairness. This has benefits on single disks and JBODs, but causes harm on raids.

I didn't test RAID as that machine with hardware RAID HBA is crashed now. But if we turn on hardware RAID in HBA, mostly we use noop io scheduler.

I think you should start testing cfq with them, too. From 2.6.33, we have some big improvements in this area.

Great! I once compared cfq and noop against non-raid and raid0. One interesting finding about sequential read testing is when there are fewer processes to read files on the raid0 JBOD, noop on raid0 is pretty good, but when there are lots of processes to do so on a non-raid JBOD, cfq is pretty better. I planed to investigate it, but too busy in other issues.

Quoted 69 lines
Quoted 58 lines
Quoted 12 lines

For 2.6.33, we changed the way in which this is handled, restoring the enable_idle = 0 for seeky queues as it was in 2.6.31:

(quoted diff) +1 -1
@@ -2218,13 +2352,10 @@ cfq_update_idle_window(struct cfq_data *cfqd,
 struct cfq_queue *cfqq,
       enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
 
       if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
-           (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
+           (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))
               enable_idle = 0;

(compare with 2.6.31:

Quoted 7 lines
        if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
            (cfqd->hw_tag && CIC_SEEKY(cic)))
                enable_idle = 0;

excluding the sample_valid check, it should be equivalent for you (I assume you have NCQ disks))
and we provide fairness for them by servicing all seeky queues together, and then idling before switching to other ones.

As for function cfq_update_idle_window, you is right. But since 2.6.32, CFQ merges many patches and the patches have impact on each other.

Quoted 36 lines

The mmap 64k randreader will have a large seek_mean, resulting in being marked seeky, but will send 16 * 4k sequential requests one after the other, so alternating between those seeky queues will cause harm.

I'm working on a new way to compute seekiness of queues, that should fix your issue, correctly identifying those queues as non-seeky (for me, a queue should be considered seeky only if it submits more than 1 seeky requests for 8 sequential ones).

Quoted 25 lines

The test scenario: 1 JBOD has 12 disks and every disk has 2 partitions. Create 8 1-GB files per partition and start 8 processes to do rand read on the 8 files per partitions. There are 8*24 processes totally. randread block size is 64K.

We found the regression on 2 machines. One machine has 8GB memory and the other has 6GB.

Bisect is very unstable. The related patches are many instead of just one.

1) commit 8e550632cccae34e265cb066691945515eaa7fb5

Author: Corrado Zoccolo <[email protected]>
Date:   Thu Nov 26 10:02:58 2009 +0100

cfq-iosched: fix corner cases in idling logic

This patch introduces about less than 20% regression. I just reverted below section and this part regression disappear. It shows this regression is stable and not impacted by other patches.

(quoted diff) +0 -0
@@ -1253,9 +1254,9 @@ static void cfq_arm_slice_timer(struct cfq_data *cfqd)
               return;
 
Quoted 10 lines
Quoted 7 lines

Review static void cfq_arm_slice_timer(struct cfq_data *cfqd) · line 1260

Quoted 7 lines
(quoted diff) +2 -2
       /*
-        * still requests with the driver, don't idle
+        * still active requests from this queue, don't idle
        */
-       if (rq_in_driver(cfqd))
+       if (cfqq->dispatched)
               return;

Although 5 patches are related to the regression, above line is quite independent. Reverting above line could always improve the result for about 20%.

I've looked at your fio script, and it is quite complex,

Quoted 8 lines
Quoted 3 lines

As we have about 40 fio sub cases, we have a script to create fio job file from a specific parameter list. So there are some superfluous parameters.

My point is that there are so many things going on, that is more difficult to analyse the issues.
I prefer looking at one problem at a time, so (initially) removing the possibility of queue merging, that Shaohua already investigated, can help in spotting the still not-well-understood problem.

Sounds reasonable.

Quoted 2 lines

Could you generate the same script, but with each process accessing only one of the files, instead of chosing it at random?

Ok. New testing starts 8 processes per partition and every process just works on one file.

Quoted 41 lines
Quoted 41 lines

Another point is we need stable result.

Quoted 3 lines

with lot of
things going on.
Let's keep this for last.

Ok. But the change like what you do mostly reduces regresion.

Quoted 33 lines

I've created a smaller test, that already shows some regression:
[global]
direct=0
ioengine=mmap
size=8G
bs=64k
numjobs=1
loops=5
runtime=60
#group_reporting
invalidate=0
directory=/media/hd/cfq-tests

[job0]
startdelay=0
rw=randread
filename=testfile1

[job1]
startdelay=0
rw=randread
filename=testfile2

[job2]
startdelay=0
rw=randread
filename=testfile3

[job3]
startdelay=0
rw=randread
filename=testfile4

Quoted 7 lines
Quoted 6 lines
Quoted 4 lines

The attached patches, in particular 0005 (that apply on top of for-linus branch of Jen's tree
git://git.kernel.dk/linux-2.6-block.git) fix the regression on this simplified workload.

I didn't download the tree. I tested the 3 attached patches against 2.6.33-rc1. The result isn't resolved.

Can you quantify if there is an improvement, though?

Ok. Because of company policy, I could only post percent instead of real number.

Quoted 6 lines

Please, also include Shahoua's patches.
I'd like to see the comparison between (always with low_latency set to 0):
plain 2.6.33
plain 2.6.33 + shahoua's
plain 2.6.33 + shahoua's + my patch
plain 2.6.33 + shahoua's + my patch + rq_in_driver vs dispatched patch.

1) low_latency=0
2.6.32 kernel 0
2.6.33-rc1 -0.33
2.6.33-rc1_shaohua -0.33
2.6.33-rc1+corrado 0.03
2.6.33-rc1_corrado+shaohua 0.02
2.6.33-rc1_corrado+shaohua+rq_in_driver 0.01

2) low_latency=1
2.6.32 kernel 0
2.6.33-rc1 -0.45
2.6.33-rc1+corrado -0.24
2.6.33-rc1_corrado+shaohua -0.23
2.6.33-rc1_corrado+shaohua+rq_in_driver -0.23

When low_latency=1, we get the biggest number with kernel 2.6.32. Comparing with low_latency=0's result, the prior one is about 4% better.

Quoted 68 lines
Quoted 65 lines
Quoted 18 lines
Quoted 6 lines
Quoted 5 lines
Quoted 4 lines

This shouldn't affect you if all queues are marked as idle.

Do you mean to use command ionice to mark it as idle class? I didn't try it.

No. I meant forcing enable_idle = 1, as you were almost doing with your patch, when cfq_latency was set.
With my above patch, this should not be needed any more, since the queues should be seen as sequential.

Quoted 7 lines
Quoted 7 lines

Does just
your patch:

Quoted 4 lines
block/cfq-iosched.c +2 -2
-           (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
-            && CFQQ_SEEKY(cfqq)))
+           (!cfqd->cfq_latency && !cfq_cfqq_deep(cfqq) &&
+               sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))
Quoted 8 lines
Quoted 5 lines
Quoted 1 line

fix most of the regression without touching arm_slice_timer?

No. If to fix the regression completely, I need apply above patch plus a debug patch. The debug patch is to just work around the 3 patches report by Shaohua's tiobench regression report. Without the debug patch, the regression isn't resolved.

Jens already merged one of Shaohua's patches, that may fix the problem with queue combining.

I did another testing. Apply my debug patch+ the low_latency patch, but use Shaohua's 2 patches (improve merge and split), the regression disappears.

Quoted 30 lines

block/cfq-iosched.c 2009-12-30 17:12:28.000000000 +0800 cfq_set_prio_slice(struct cfq_data *cfqd · line 597

Quoted 12 lines

Below is the debug patch.
diff -Nraup linux-2.6.33_rc1/block/cfq-iosched.c linux-2.6.33_rc1_randread64k/block/cfq-iosched.c

block/cfq-iosched.c 2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c 2009-12-23 14:12:03.000000000 +0800 +3 -0
@@ -592,6 +592,9 @@ cfq_set_prio_slice(struct cfq_data *cfqd
       cfqq->slice_start = jiffies;
       cfqq->slice_end = jiffies + slice;
       cfqq->allocated_slice = slice;
+/*YMZHANG*/
+       cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
+

This is disabled, on a vanilla 2.6.33 kernel, by setting low_latency = 0

Quoted 17 lines
block/cfq-iosched.c 2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c 2009-12-23 14:12:03.000000000 +0800 +2 -1
       cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
 }
 
@@ -1836,7 +1839,8 @@ static void cfq_arm_slice_timer(struct c
       /*
        * still active requests from this queue, don't idle
        */
-       if (cfqq->dispatched)
+       //if (cfqq->dispatched)
+       if (rq_in_driver(cfqd))
               return;
 
       /*
@@ -1941,6 +1945,9 @@ static void cfq_setup_merge(struct cfq_q
               new_cfqq = __cfqq;
       }
 
Quoted 6 lines

block/cfq-iosched.c 2009-12-30 17:12:28.000000000 +0800 static void cfq_setup_merge(struct cfq_q · line 1950

Quoted 3 lines
block/cfq-iosched.c 2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c 2009-12-23 14:12:03.000000000 +0800 +3 -0
+       /* YMZHANG debug */
+       return;
+

This should be partially addressed by Shaohua's patch merged in Jens' tree. But note that your 8 processes, can randomly start doing I/O on the same file, so merging those queues is sometimes reasonable.

Quoted 3 lines

Another reason is I start 8 processes per partition and every disk has 2 partitions, so there are 16 processes per disk. With another JBOD, I use one partition per disk, and the regression is only 8%.

Quoted 6 lines

With half of the processes, time slices are higher, and the disk cache can do a better job when servicing interleaved sequential requests.

Quoted 2 lines
Quoted 2 lines

From this point, can CFQ do not merge request queues which access different partitions?

(puzzled: I didn't write this, and can't find a message in the thread with this question.)

My email client is evolution and sometimes it adds > unexpectedly.

Quoted 5 lines
Quoted 2 lines

As you know, it's unusual that a process accesses files across partitions. io scheduler is at low layer which doesn't know partition.

CFQ bases decision on distance between requests, and requests going to different partitions will have much higher distance. So the associated queues will be more likely marked as seeky.

Right. Thanks for your explanation.

Quoted 40 lines
Quoted 40 lines
Quoted 40 lines

The patch to split them quickly was still not merged, though, so you will still see some regression due to this. In my simplified job file, I removed the randomness to make sure this cannot happen.

Quoted 8 lines
block/cfq-iosched.c 2009-12-30 17:12:28.000000000 +0800 renamed from block/cfq-iosched.c 2009-12-23 14:12:03.000000000 +0800 +0 -0
       process_refs = cfqq_process_refs(cfqq);
       /*
        * If the process for the cfqq has gone away, there is no
Quoted 5 lines

I guess

Quoted 1 line

5db5d64277bf390056b1a87d0bb288c8b8553f96.

Quoted 8 lines
Quoted 2 lines

will still introduce a 10% regression, but this is needed to improve latency, and you can just disable low_latency to avoid it.

You are right. I did a quick testing. If my patch + revert 2 patches and keep 5db5d64, the regression is about 20%.

But low_latency=0 doesn't work like what we imagined. If patch + revert 2 patches and keep 5db5d64 while set low_latency=0, the regression is still there. One reason is my patch doesn't work when low_latency=0.

Right. You can try with my patch, instead, that doesn't depend on low_latency, and set it to 0 to remove this performance degradation. My results:
2.6.32.2:

   READ: io=146688KB, aggrb=2442KB/s, minb=602KB/s, maxb=639KB/s,

mint=60019msec, maxt=60067msec

2.6.33 - jens:

   READ: io=128512KB, aggrb=2140KB/s, minb=526KB/s, maxb=569KB/s,

mint=60004msec, maxt=60032msec

2.6.33 - jens + my patches :

   READ: io=143232KB, aggrb=2384KB/s, minb=595KB/s, maxb=624KB/s,

mint=60003msec, maxt=60072msec

2.6.33 - jens + my patches + low_lat = 0:

   READ: io=145216KB, aggrb=2416KB/s, minb=596KB/s, maxb=632KB/s,

mint=60027msec, maxt=60087msec

#7 lore raw

Hi Yanmin,

Quoted 114 lines · Yanmin
Quoted 105 lines · Corrado Zoccolo

Hi

Quoted 102 lines · Yanmin
Quoted 19 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 16 lines · Yanmin
Quoted 6 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines · Yanmin

Comparing with kernel 2.6.32, fio mmap randread 64k has more than 40% regression with 2.6.33-rc1.

Thanks for your timely reply. Some comments inlined below.

Quoted 1 line

Can you compare the performance also with 2.6.31?

We did. We run Linux kernel Performance Tracking project and run many benchmarks when a RC kernel is released.

The result of 2.6.31 is quite similar to the one of 2.6.32. But the one of 2.6.30 is about 8% better than the one of 2.6.31.

Quoted 8 lines
Quoted 6 lines
Quoted 4 lines

I think I understand what causes your problem.
2.6.32, with default settings, handled even random readers as sequential ones to provide fairness. This has benefits on single disks and JBODs, but causes harm on raids.

I didn't test RAID as that machine with hardware RAID HBA is crashed now. But if we turn on hardware RAID in HBA, mostly we use noop io scheduler.

I think you should start testing cfq with them, too. From 2.6.33, we have some big improvements in this area.

Great! I once compared cfq and noop against non-raid and raid0. One interesting finding about sequential read testing is when there are fewer processes to read files on the raid0 JBOD, noop on raid0 is pretty good, but when there are lots of processes to do so on a non-raid JBOD, cfq is pretty better. I planed to investigate it, but too busy in other issues.

Quoted 69 lines
Quoted 58 lines
Quoted 11 lines

For 2.6.33, we changed the way in which this is handled, restoring the enable_idle = 0 for seeky queues as it was in 2.6.31:

(quoted diff) +0 -0
@@ -2218,13 +2352,10 @@ cfq_update_idle_window(struct cfq_data *cfqd,
 struct cfq_queue *cfqq,
        enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
 

       if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
-           (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
+           (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))) enable_idle = 0;

Quoted 8 lines

(compare with 2.6.31:
        if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || (cfqd->hw_tag && CIC_SEEKY(cic)))
                enable_idle = 0;
excluding the sample_valid check, it should be equivalent for you (I assume you have NCQ disks))
and we provide fairness for them by servicing all seeky queues together, and then idling before switching to other ones.

As for function cfq_update_idle_window, you is right. But since 2.6.32, CFQ merges many patches and the patches have impact on each other.

Quoted 36 lines

The mmap 64k randreader will have a large seek_mean, resulting in being marked seeky, but will send 16 * 4k sequential requests one after the other, so alternating between those seeky queues will cause harm.

I'm working on a new way to compute seekiness of queues, that should fix your issue, correctly identifying those queues as non-seeky (for me, a queue should be considered seeky only if it submits more than 1 seeky requests for 8 sequential ones).

Quoted 25 lines

The test scenario: 1 JBOD has 12 disks and every disk has 2 partitions. Create 8 1-GB files per partition and start 8 processes to do rand read on the 8 files per partitions. There are 8*24 processes totally. randread block size is 64K.

We found the regression on 2 machines. One machine has 8GB memory and the other has 6GB.

Bisect is very unstable. The related patches are many instead of just one.

1) commit 8e550632cccae34e265cb066691945515eaa7fb5
Author: Corrado Zoccolo <[email protected]>
Date:   Thu Nov 26 10:02:58 2009 +0100

   cfq-iosched: fix corner cases in idling logic

This patch introduces about less than 20% regression. I just reverted below section and this part regression disappear. It shows this regression is stable and not impacted by other patches.

(quoted diff) +0 -0
@@ -1253,9 +1254,9 @@ static void cfq_arm_slice_timer(struct cfq_data *cfqd)
                return;
 
Quoted 10 lines
Quoted 7 lines

Review static void cfq_arm_slice_timer(struct cfq_data *cfqd) · line 1260

Quoted 7 lines
(quoted diff) +2 -2
        /*
-        * still requests with the driver, don't idle
+        * still active requests from this queue, don't idle
         */
-       if (rq_in_driver(cfqd))
+       if (cfqq->dispatched)
                return;

Although 5 patches are related to the regression, above line is quite independent. Reverting above line could always improve the result for about 20%.

I've looked at your fio script, and it is quite complex,

Quoted 8 lines
Quoted 3 lines

As we have about 40 fio sub cases, we have a script to create fio job file from a specific parameter list. So there are some superfluous parameters.

My point is that there are so many things going on, that is more difficult to analyse the issues.
I prefer looking at one problem at a time, so (initially) removing the possibility of queue merging, that Shaohua already investigated, can help in spotting the still not-well-understood problem.

Quoted 6 lines

Sounds reasonable.

Quoted 2 lines

Could you generate the same script, but with each process accessing only one of the files, instead of chosing it at random?

Ok. New testing starts 8 processes per partition and every process just works on one file.

Great, thanks.

Quoted 46 lines
Quoted 46 lines
Quoted 46 lines

Another point is we need stable result.

Quoted 3 lines

with lot of
things going on.
Let's keep this for last.

Ok. But the change like what you do mostly reduces regresion.

Quoted 37 lines

I've created a smaller test, that already shows some regression:
[global]
direct=0
ioengine=mmap
size=8G
bs=64k
numjobs=1
loops=5
runtime=60
#group_reporting
invalidate=0
directory=/media/hd/cfq-tests

[job0]
startdelay=0
rw=randread
filename=testfile1

[job1]
startdelay=0
rw=randread
filename=testfile2

[job2]
startdelay=0
rw=randread
filename=testfile3

[job3]
startdelay=0
rw=randread
filename=testfile4

The attached patches, in particular 0005 (that apply on top of for-linus branch of Jen's tree
git://git.kernel.dk/linux-2.6-block.git) fix the regression on this simplified workload.

Quoted 5 lines
Quoted 3 lines
Quoted 2 lines

I didn't download the tree. I tested the 3 attached patches against 2.6.33-rc1. The result isn't resolved.

Can you quantify if there is an improvement, though?

Ok. Because of company policy, I could only post percent instead of real number

Sure, it is fine.

Quoted 8 lines
Quoted 7 lines

Please, also include Shahoua's patches.
I'd like to see the comparison between (always with low_latency set to 0):
plain 2.6.33
plain 2.6.33 + shahoua's
plain 2.6.33 + shahoua's + my patch
plain 2.6.33 + shahoua's + my patch + rq_in_driver vs dispatched patch.

Quoted 8 lines

1) low_latency=0
2.6.32 kernel                                   0 2.6.33-rc1                                      -0.33 2.6.33-rc1_shaohua                              -0.33 2.6.33-rc1+corrado                              0.03 2.6.33-rc1_corrado+shaohua                      0.02 2.6.33-rc1_corrado+shaohua+rq_in_driver         0.01

So my patch fixes the situation for low_latency = 0, as I expected. I'll send it to Jens with a proper changelog.

Quoted 6 lines

2) low_latency=1
2.6.32 kernel                                   0 2.6.33-rc1                                      -0.45 2.6.33-rc1+corrado                              -0.24 2.6.33-rc1_corrado+shaohua                      -0.23 2.6.33-rc1_corrado+shaohua+rq_in_driver         -0.23

The results are as expected. With each process working on a separate file, Shahoua's patches do not influence the result sensibly. Interestingly, even rq_in_driver doesn't improve in this case, so maybe its effect is somewhat connected to queue merging. The remaining -23% is due to timeslice shrinking, that is done to reduce max latency when there are too many processes doing I/O, at the expense of throughput. It is a documented change, and the suggested way if you favor throughput over latency is to set low_latency = 0.

Quoted 4 lines

When low_latency=1, we get the biggest number with kernel 2.6.32. Comparing with low_latency=0's result, the prior one is about 4% better.

Ok, so 2.6.33 + corrado (with low_latency =0) is comparable with fastest 2.6.32, so we can consider the first part of the problem solved.

For the queue merging issue, maybe Jeff has some improvements w.r.t shaohua's approach.

Thanks,
Corrado

#8 lore raw

Hi Yanmin

Quoted 212 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 114 lines · Yanmin
Quoted 105 lines · Corrado Zoccolo

Hi

Quoted 102 lines · Yanmin
Quoted 19 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 16 lines · Yanmin
Quoted 6 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines · Yanmin

Comparing with kernel 2.6.32, fio mmap randread 64k has more than 40% regression with 2.6.33-rc1.

Thanks for your timely reply. Some comments inlined below.

Quoted 1 line

Can you compare the performance also with 2.6.31?

We did. We run Linux kernel Performance Tracking project and run many benchmarks when a RC kernel is released.

The result of 2.6.31 is quite similar to the one of 2.6.32. But the one of 2.6.30 is about 8% better than the one of 2.6.31.

Quoted 8 lines
Quoted 6 lines
Quoted 4 lines

I think I understand what causes your problem.
2.6.32, with default settings, handled even random readers as sequential ones to provide fairness. This has benefits on single disks and JBODs, but causes harm on raids.

I didn't test RAID as that machine with hardware RAID HBA is crashed now. But if we turn on hardware RAID in HBA, mostly we use noop io scheduler.

I think you should start testing cfq with them, too. From 2.6.33, we have some big improvements in this area.

Great! I once compared cfq and noop against non-raid and raid0. One interesting finding about sequential read testing is when there are fewer processes to read files on the raid0 JBOD, noop on raid0 is pretty good, but when there are lots of processes to do so on a non-raid JBOD, cfq is pretty better. I planed to investigate it, but too busy in other issues.

Quoted 69 lines
Quoted 58 lines
Quoted 11 lines

For 2.6.33, we changed the way in which this is handled, restoring the enable_idle = 0 for seeky queues as it was in 2.6.31:

(quoted diff) +0 -0
@@ -2218,13 +2352,10 @@ cfq_update_idle_window(struct cfq_data *cfqd,
 struct cfq_queue *cfqq,
        enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
 

       if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
-           (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
+           (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))) enable_idle = 0;

Quoted 8 lines

(compare with 2.6.31:
        if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || (cfqd->hw_tag && CIC_SEEKY(cic)))
                enable_idle = 0;
excluding the sample_valid check, it should be equivalent for you (I assume you have NCQ disks))
and we provide fairness for them by servicing all seeky queues together, and then idling before switching to other ones.

As for function cfq_update_idle_window, you is right. But since 2.6.32, CFQ merges many patches and the patches have impact on each other.

Quoted 36 lines

The mmap 64k randreader will have a large seek_mean, resulting in being marked seeky, but will send 16 * 4k sequential requests one after the other, so alternating between those seeky queues will cause harm.

I'm working on a new way to compute seekiness of queues, that should fix your issue, correctly identifying those queues as non-seeky (for me, a queue should be considered seeky only if it submits more than 1 seeky requests for 8 sequential ones).

Quoted 25 lines

The test scenario: 1 JBOD has 12 disks and every disk has 2 partitions. Create 8 1-GB files per partition and start 8 processes to do rand read on the 8 files per partitions. There are 8*24 processes totally. randread block size is 64K.

We found the regression on 2 machines. One machine has 8GB memory and the other has 6GB.

Bisect is very unstable. The related patches are many instead of just one.

1) commit 8e550632cccae34e265cb066691945515eaa7fb5
Author: Corrado Zoccolo <[email protected]>
Date:   Thu Nov 26 10:02:58 2009 +0100

   cfq-iosched: fix corner cases in idling logic

This patch introduces about less than 20% regression. I just reverted below section and this part regression disappear. It shows this regression is stable and not impacted by other patches.

(quoted diff) +0 -0
@@ -1253,9 +1254,9 @@ static void cfq_arm_slice_timer(struct cfq_data *cfqd)
                return;
 
Quoted 10 lines
Quoted 7 lines

Review static void cfq_arm_slice_timer(struct cfq_data *cfqd) · line 1260

Quoted 7 lines
(quoted diff) +2 -2
        /*
-        * still requests with the driver, don't idle
+        * still active requests from this queue, don't idle
         */
-       if (rq_in_driver(cfqd))
+       if (cfqq->dispatched)
                return;

Although 5 patches are related to the regression, above line is quite independent. Reverting above line could always improve the result for about 20%.

I've looked at your fio script, and it is quite complex,

Quoted 8 lines
Quoted 3 lines

As we have about 40 fio sub cases, we have a script to create fio job file from a specific parameter list. So there are some superfluous parameters.

My point is that there are so many things going on, that is more difficult to analyse the issues.
I prefer looking at one problem at a time, so (initially) removing the possibility of queue merging, that Shaohua already investigated, can help in spotting the still not-well-understood problem.

Quoted 6 lines

Sounds reasonable.

Quoted 2 lines

Could you generate the same script, but with each process accessing only one of the files, instead of chosing it at random?

Ok. New testing starts 8 processes per partition and every process just works on one file.

Great, thanks.

Quoted 46 lines
Quoted 46 lines
Quoted 46 lines

Another point is we need stable result.

Quoted 3 lines

with lot of
things going on.
Let's keep this for last.

Ok. But the change like what you do mostly reduces regresion.

Quoted 37 lines

I've created a smaller test, that already shows some regression:
[global]
direct=0
ioengine=mmap
size=8G
bs=64k
numjobs=1
loops=5
runtime=60
#group_reporting
invalidate=0
directory=/media/hd/cfq-tests

[job0]
startdelay=0
rw=randread
filename=testfile1

[job1]
startdelay=0
rw=randread
filename=testfile2

[job2]
startdelay=0
rw=randread
filename=testfile3

[job3]
startdelay=0
rw=randread
filename=testfile4

The attached patches, in particular 0005 (that apply on top of for-linus branch of Jen's tree
git://git.kernel.dk/linux-2.6-block.git) fix the regression on this simplified workload.

Quoted 5 lines
Quoted 3 lines
Quoted 2 lines

I didn't download the tree. I tested the 3 attached patches against 2.6.33-rc1. The result isn't resolved.

Can you quantify if there is an improvement, though?

Ok. Because of company policy, I could only post percent instead of real number

Sure, it is fine.

Quoted 8 lines
Quoted 7 lines

Please, also include Shahoua's patches.
I'd like to see the comparison between (always with low_latency set to 0):
plain 2.6.33
plain 2.6.33 + shahoua's
plain 2.6.33 + shahoua's + my patch
plain 2.6.33 + shahoua's + my patch + rq_in_driver vs dispatched patch.

Quoted 8 lines

1) low_latency=0
2.6.32 kernel                                   0 2.6.33-rc1                                      -0.33 2.6.33-rc1_shaohua                              -0.33 2.6.33-rc1+corrado                              0.03 2.6.33-rc1_corrado+shaohua                      0.02 2.6.33-rc1_corrado+shaohua+rq_in_driver         0.01

So my patch fixes the situation for low_latency = 0, as I expected. I'll send it to Jens with a proper changelog.

Quoted 6 lines

2) low_latency=1
2.6.32 kernel                                   0 2.6.33-rc1                                      -0.45 2.6.33-rc1+corrado                              -0.24 2.6.33-rc1_corrado+shaohua                      -0.23 2.6.33-rc1_corrado+shaohua+rq_in_driver         -0.23

The results are as expected. With each process working on a separate file, Shahoua's patches do not influence the result sensibly. Interestingly, even rq_in_driver doesn't improve in this case, so maybe its effect is somewhat connected to queue merging. The remaining -23% is due to timeslice shrinking, that is done to reduce max latency when there are too many processes doing I/O, at the expense of throughput. It is a documented change, and the suggested way if you favor throughput over latency is to set low_latency = 0.

Quoted 2 lines
Quoted 6 lines
Quoted 2 lines

When low_latency=1, we get the biggest number with kernel 2.6.32. Comparing with low_latency=0's result, the prior one is about 4% better.

Ok, so 2.6.33 + corrado (with low_latency =0) is comparable with fastest 2.6.32, so we can consider the first part of the problem solved.

I think we can return now to your full script with queue merging. I'm wondering if (in arm_slice_timer):
- if (cfqq->dispatched)
+ if (cfqq->dispatched || (cfqq->new_cfqq && rq_in_driver(cfqd)))

               return;

gives the same improvement you were experiencing just reverting to rq_in_driver.

We saw that cfqq->dispatched worked fine when there was no queue merging happening, so it must be something concerning merging, probably dispatched is not accurate when we set up for a merging, but the merging was not yet done.

Thanks,
Corrado

#9 lore raw
Quoted 14 lines · Corrado Zoccolo

Hi Yanmin

Quoted 7 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines

When low_latency=1, we get the biggest number with kernel 2.6.32. Comparing with low_latency=0's result, the prior one is about 4% better.

Ok, so 2.6.33 + corrado (with low_latency =0) is comparable with fastest 2.6.32, so we can consider the first part of the problem solved.

I think we can return now to your full script with queue merging. I'm wondering if (in arm_slice_timer):

(quoted diff) +1 -1
-       if (cfqq->dispatched)
+      if (cfqq->dispatched || (cfqq->new_cfqq && rq_in_driver(cfqd)))
               return;
Quoted 1 line

gives the same improvement you were experiencing just reverting to rq_in_driver.

I did a quick testing against 2.6.33-rc1. With the new method, fio mmap randread 46k has about 20% improvement. With just checking rq_in_driver(cfqd), it has about 33% improvement.

Quoted 5 lines

We saw that cfqq->dispatched worked fine when there was no queue merging happening, so it must be something concerning merging, probably dispatched is not accurate when we set up for a merging, but the merging was not yet done.

#10 lore raw
Quoted 16 lines · Yanmin
Quoted 9 lines · Corrado Zoccolo

Hi Yanmin

Quoted 7 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines

When low_latency=1, we get the biggest number with kernel 2.6.32. Comparing with low_latency=0's result, the prior one is about 4% better.

Ok, so 2.6.33 + corrado (with low_latency =0) is comparable with fastest 2.6.32, so we can consider the first part of the problem solved.

Quoted 6 lines

I think we can return now to your full script with queue merging. I'm wondering if (in arm_slice_timer):
-       if (cfqq->dispatched)
+      if (cfqq->dispatched || (cfqq->new_cfqq && rq_in_driver(cfqd))) return;
gives the same improvement you were experiencing just reverting to rq_in_driver.

Quoted 4 lines

I did a quick testing against 2.6.33-rc1. With the new method, fio mmap randread 46k has about 20% improvement. With just checking rq_in_driver(cfqd), it has about 33% improvement.

Jeff, do you have an idea why in arm_slice_timer, checking rq_in_driver instead of cfqq->dispatched gives so much improvement in presence of queue merging, while it doesn't have noticeable effect when there are no merges?

Thanks,
Corrado

Quoted 9 lines
Quoted 6 lines

We saw that cfqq->dispatched worked fine when there was no queue merging happening, so it must be something concerning merging, probably dispatched is not accurate when we set up for a merging, but the merging was not yet done.

#11 lore raw
Quoted 18 lines · Corrado Zoccolo
Quoted 16 lines · Yanmin
Quoted 15 lines · Corrado Zoccolo

Hi Yanmin

Quoted 7 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines

When low_latency=1, we get the biggest number with kernel 2.6.32. Comparing with low_latency=0's result, the prior one is about 4% better.

Ok, so 2.6.33 + corrado (with low_latency =0) is comparable with fastest 2.6.32, so we can consider the first part of the problem solved.

I think we can return now to your full script with queue merging. I'm wondering if (in arm_slice_timer):
-       if (cfqq->dispatched)
+      if (cfqq->dispatched || (cfqq->new_cfqq && rq_in_driver(cfqd))) return;
gives the same improvement you were experiencing just reverting to rq_in_driver.

Quoted 8 lines
Quoted 4 lines

I did a quick testing against 2.6.33-rc1. With the new method, fio mmap randread 46k has about 20% improvement. With just checking rq_in_driver(cfqd), it has about 33% improvement.

Jeff, do you have an idea why in arm_slice_timer, checking rq_in_driver instead of cfqq->dispatched gives so much improvement in presence of queue merging, while it doesn't have noticeable effect when there are no merges?

It's tough to say. Is there any chance I could get some blktrace data for the run?

Cheers,
Jeff

#12 lore raw
In reply to Corrado Zoccolo
Quoted 18 lines · Corrado Zoccolo
Quoted 16 lines · Yanmin
Quoted 15 lines · Corrado Zoccolo

Hi Yanmin

Quoted 7 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines

When low_latency=1, we get the biggest number with kernel 2.6.32. Comparing with low_latency=0's result, the prior one is about 4% better.

Ok, so 2.6.33 + corrado (with low_latency =0) is comparable with fastest 2.6.32, so we can consider the first part of the problem solved.

I think we can return now to your full script with queue merging. I'm wondering if (in arm_slice_timer):
-       if (cfqq->dispatched)
+      if (cfqq->dispatched || (cfqq->new_cfqq && rq_in_driver(cfqd))) return;
gives the same improvement you were experiencing just reverting to rq_in_driver.

Quoted 8 lines
Quoted 4 lines

I did a quick testing against 2.6.33-rc1. With the new method, fio mmap randread 46k has about 20% improvement. With just checking rq_in_driver(cfqd), it has about 33% improvement.

Jeff, do you have an idea why in arm_slice_timer, checking rq_in_driver instead of cfqq->dispatched gives so much improvement in presence of queue merging, while it doesn't have noticeable effect when there are no merges?

Performance improvement because of replacing cfqq->dispatched with
rq_in_driver() is really strange. This will mean we will do even lesser
idling on the cfqq. That means faster cfqq switching and that should mean more
seeks (for this test case) and reduce throughput. This is just opposite to your approach of treating a random read mmap queue as sync where we will idle on the queue.

Thanks
Vivek

Quoted 13 lines

Thanks,
Corrado

Quoted 9 lines
Quoted 6 lines

We saw that cfqq->dispatched worked fine when there was no queue merging happening, so it must be something concerning merging, probably dispatched is not accurate when we set up for a merging, but the merging was not yet done.

#13 lore raw
Quoted 28 lines · Vivek Goyal
Quoted 18 lines · Corrado Zoccolo
Quoted 16 lines · Yanmin
Quoted 15 lines · Corrado Zoccolo

Hi Yanmin

Quoted 7 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines

When low_latency=1, we get the biggest number with kernel 2.6.32. Comparing with low_latency=0's result, the prior one is about 4% better.

Ok, so 2.6.33 + corrado (with low_latency =0) is comparable with fastest 2.6.32, so we can consider the first part of the problem solved.

I think we can return now to your full script with queue merging. I'm wondering if (in arm_slice_timer):
-       if (cfqq->dispatched)
+      if (cfqq->dispatched || (cfqq->new_cfqq && rq_in_driver(cfqd))) return;
gives the same improvement you were experiencing just reverting to rq_in_driver.

Quoted 8 lines
Quoted 4 lines

I did a quick testing against 2.6.33-rc1. With the new method, fio mmap randread 46k has about 20% improvement. With just checking rq_in_driver(cfqd), it has about 33% improvement.

Jeff, do you have an idea why in arm_slice_timer, checking rq_in_driver instead of cfqq->dispatched gives so much improvement in presence of queue merging, while it doesn't have noticeable effect when there are no merges?

Quoted 5 lines

Performance improvement because of replacing cfqq->dispatched with
rq_in_driver() is really strange. This will mean we will do even lesser
idling on the cfqq. That means faster cfqq switching and that should mean more
seeks (for this test case) and reduce throughput. This is just opposite to your approach of treating a random read mmap queue as sync where we will idle on the queue.

The tests (previous mails in this thread) show that, if no queue merging is happening, handling the queue as sync_idle, and setting low_latency = 0 to have bigger slices completely recovers the regression.
If, though, we have queue merges, current arm_slice_timer shows regression w.r.t. the rq_in_driver version (2.6.32).
I think a possible explanation is that we are idling instead of switching to an other queue that would be merged with this one. In fact, my half-backed try to have the rq_in_driver check conditional on queue merging fixed part of the regression (not all, because queue merges are not symmetrical, and I could be seeing the queue that is 'new_cfqq' for an other).

Thanks,
Corrado

Quoted 18 lines

Thanks
Vivek

Quoted 13 lines

Thanks,
Corrado

Quoted 9 lines
Quoted 6 lines

We saw that cfqq->dispatched worked fine when there was no queue merging happening, so it must be something concerning merging, probably dispatched is not accurate when we set up for a merging, but the merging was not yet done.

Signature · 10 lines
__________________________________________________________________________

dott. Corrado Zoccolo                          mailto:[email protected]
PhD - Department of Computer Science - University of Pisa, Italy
--------------------------------------------------------------------------
The self-confidence of a warrior is not the self-confidence of the average
man. The average man seeks certainty in the eyes of the onlooker and calls
that self-confidence. The warrior seeks impeccability in his own eyes and
calls that humbleness.
                               Tales of Power - C. Castaneda
#15 lore raw
In reply to Vivek Goyal
Quoted 28 lines · Vivek Goyal
Quoted 18 lines · Corrado Zoccolo
Quoted 16 lines · Yanmin
Quoted 15 lines · Corrado Zoccolo

Hi Yanmin

Quoted 7 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines

When low_latency=1, we get the biggest number with kernel 2.6.32. Comparing with low_latency=0's result, the prior one is about 4% better.

Ok, so 2.6.33 + corrado (with low_latency =0) is comparable with fastest 2.6.32, so we can consider the first part of the problem solved.

I think we can return now to your full script with queue merging. I'm wondering if (in arm_slice_timer):

(quoted diff) +1 -1
-       if (cfqq->dispatched)
+      if (cfqq->dispatched || (cfqq->new_cfqq && rq_in_driver(cfqd)))
               return;

gives the same improvement you were experiencing just reverting to rq_in_driver.

Quoted 8 lines
Quoted 4 lines

I did a quick testing against 2.6.33-rc1. With the new method, fio mmap randread 46k has about 20% improvement. With just checking rq_in_driver(cfqd), it has about 33% improvement.

Jeff, do you have an idea why in arm_slice_timer, checking rq_in_driver instead of cfqq->dispatched gives so much improvement in presence of queue merging, while it doesn't have noticeable effect when there are no merges?

Quoted 5 lines

Performance improvement because of replacing cfqq->dispatched with
rq_in_driver() is really strange. This will mean we will do even lesser
idling on the cfqq. That means faster cfqq switching and that should mean more
seeks (for this test case) and reduce throughput. This is just opposite to your approach of treating a random read mmap queue as sync where we will idle on the queue.

I used to look at the issue, but not fully understand it. Some interesting finding:
the cfqq->dispatched cause cfq_select_queue frequently switch queues. it appears frequent switch can make we could quickly switch to sequential requests in the workload. without the cfqq->dispatched, we dispatch queue1 request, M requests from other queues, queue1 request. with it, we dispatch queue1 request, N requests from other queues, queue1 request. It appears M < N from blktrace, which cause we have less seeky. I don't see any other obvious difference from blktrace in the two cases.

Thanks,
Shaohua

#16 lore raw
Quoted 34 lines · Shaohua Li
Quoted 28 lines · Vivek Goyal
Quoted 18 lines · Corrado Zoccolo
Quoted 16 lines · Yanmin
Quoted 15 lines · Corrado Zoccolo

Hi Yanmin

Quoted 7 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines

When low_latency=1, we get the biggest number with kernel 2.6.32. Comparing with low_latency=0's result, the prior one is about 4% better.

Ok, so 2.6.33 + corrado (with low_latency =0) is comparable with fastest 2.6.32, so we can consider the first part of the problem solved.

I think we can return now to your full script with queue merging. I'm wondering if (in arm_slice_timer):

(quoted diff) +1 -1
-       if (cfqq->dispatched)
+      if (cfqq->dispatched || (cfqq->new_cfqq && rq_in_driver(cfqd)))
               return;

gives the same improvement you were experiencing just reverting to rq_in_driver.

Quoted 8 lines
Quoted 4 lines

I did a quick testing against 2.6.33-rc1. With the new method, fio mmap randread 46k has about 20% improvement. With just checking rq_in_driver(cfqd), it has about 33% improvement.

Jeff, do you have an idea why in arm_slice_timer, checking rq_in_driver instead of cfqq->dispatched gives so much improvement in presence of queue merging, while it doesn't have noticeable effect when there are no merges?

Quoted 5 lines

Performance improvement because of replacing cfqq->dispatched with
rq_in_driver() is really strange. This will mean we will do even lesser
idling on the cfqq. That means faster cfqq switching and that should mean more
seeks (for this test case) and reduce throughput. This is just opposite to your approach of treating a random read mmap queue as sync where we will idle on the queue.

Quoted 10 lines

I used to look at the issue, but not fully understand it. Some interesting finding:
the cfqq->dispatched cause cfq_select_queue frequently switch queues. it appears frequent switch can make we could quickly switch to sequential requests in the workload. without the cfqq->dispatched, we dispatch queue1 request, M requests from other queues, queue1 request. with it, we dispatch queue1 request, N requests from other queues, queue1 request. It appears M < N from blktrace, which cause we have less seeky. I don't see any other obvious difference from blktrace in the two cases.

I thought there was merging and/or unmerging activity. You don't mention that here.

I'll see if I can reproduce it.

Cheers,
Jeff

#14 lore raw
In reply to Corrado Zoccolo
Quoted 34 lines · Corrado Zoccolo
Quoted 28 lines · Vivek Goyal
Quoted 18 lines · Corrado Zoccolo
Quoted 16 lines · Yanmin
Quoted 15 lines · Corrado Zoccolo

Hi Yanmin

Quoted 7 lines · Corrado Zoccolo

Hi Yanmin,

Quoted 2 lines

When low_latency=1, we get the biggest number with kernel 2.6.32. Comparing with low_latency=0's result, the prior one is about 4% better.

Ok, so 2.6.33 + corrado (with low_latency =0) is comparable with fastest 2.6.32, so we can consider the first part of the problem solved.

I think we can return now to your full script with queue merging. I'm wondering if (in arm_slice_timer):
-       if (cfqq->dispatched)
+      if (cfqq->dispatched || (cfqq->new_cfqq && rq_in_driver(cfqd))) return;
gives the same improvement you were experiencing just reverting to rq_in_driver.

Quoted 8 lines
Quoted 4 lines

I did a quick testing against 2.6.33-rc1. With the new method, fio mmap randread 46k has about 20% improvement. With just checking rq_in_driver(cfqd), it has about 33% improvement.

Jeff, do you have an idea why in arm_slice_timer, checking rq_in_driver instead of cfqq->dispatched gives so much improvement in presence of queue merging, while it doesn't have noticeable effect when there are no merges?

Quoted 5 lines

Performance improvement because of replacing cfqq->dispatched with
rq_in_driver() is really strange. This will mean we will do even lesser
idling on the cfqq. That means faster cfqq switching and that should mean more
seeks (for this test case) and reduce throughput. This is just opposite to your approach of treating a random read mmap queue as sync where we will idle on the queue.

Quoted 13 lines

The tests (previous mails in this thread) show that, if no queue merging is happening, handling the queue as sync_idle, and setting low_latency = 0 to have bigger slices completely recovers the regression.
If, though, we have queue merges, current arm_slice_timer shows regression w.r.t. the rq_in_driver version (2.6.32).
I think a possible explanation is that we are idling instead of switching to an other queue that would be merged with this one. In fact, my half-backed try to have the rq_in_driver check conditional on queue merging fixed part of the regression (not all, because queue merges are not symmetrical, and I could be seeing the queue that is 'new_cfqq' for an other).

Just a data point. I ran 8 fio mmap jobs, bs=64K, direct=1, size=2G runtime=30 with vanilla kernel (2.6.33-rc4) and with modified kernel which replaced cfqq->dispatched with rq_in_driver(cfqd).

I did not see any significant throughput improvement but I did see max_clat halfed in modified kernel.

Vanilla kernel
==============
read bw: 3701KB/s
max clat: 401050 us
Number of times idle timer was armed: 20980
Number of cfqq expired/switched: 6377
cfqq merge operations: 0

Modified kernel (rq_in_driver(cfqd))
===================================
read bw: 3645KB/s
max clat: 800515 us
Number of times idle timer was armed: 2875
Number of cfqq expired/switched: 17750
cfqq merge operations: 0

This kind of confirms that rq_in_driver(cfqd) will reduce the number of times we idle on queues and will make queue switching faster. That also explains the reduce max clat.

If that's the case, then it should also have increased the number of seeks (at least on yanmin's setup of JBOD), and reduce throughput. But instead reverse seems to be happening in his setup.

Yanmin, as Jeff mentioned, if you can capture the blktrace of vanilla and modified kernel and upload somewhere to look at, it might help.

Thanks
Vivek