Skip to contentSkip to conversation

fio rand read/write regression with 2.6.32-rc3

DISCUSSION Zhang, Yanmin 3 participants 7 messages Oct 10, 2009 – Oct 12, 2009

lore
Share & tools
lore
#0 lore raw

Comparing with 2.6.23-rc1's result, fio rand read write has regression on my 2*4 core stoakley machine (8GB memory) with a JBOD of 12 disks.

Every disk has 8 1-GB files. Start 8 sub-processes per disk and every process random chooses a file on the disk to do 36 times of file read or write on the file and then choose another file.

fio_mmap_rand_read_4k regresion is about 35%. Bisect down to patch:
1d2235152dc745c6d94bedb550fea84cffdbf768 is first bad commit commit 1d2235152dc745c6d94bedb550fea84cffdbf768

Author: Jens Axboe <[email protected]>
Date:   Fri Oct 2 19:27:04 2009 +0200

cfq-iosched: add a knob for desktop interactiveness

After I revert the patch against 2.6.23-rc3, fio_mmap_rand_read_4k regression disappears.

fio_mmap_randrw_4k has less than 20% regression and reverting the patch could restore performance.

On another stoakley machine with another JBOD, we see the similiar regression.

ffsb rand read/write has a bigger regression and reverting the patch could restore performance.

Yanmin

#1 lore raw
Quoted 9 lines · Yanmin

Comparing with 2.6.23-rc1's result, fio rand read write has regression on my 2*4 core stoakley machine (8GB memory) with a JBOD of 12 disks.

Every disk has 8 1-GB files. Start 8 sub-processes per disk and every process random chooses a file on the disk to do 36 times of file read or write on the file and then choose another file.

fio_mmap_rand_read_4k regresion is about 35%.

Heh, I seem to recollect I told Linus that this would cost is 30-40% performance. So not totally crazy.

So yes, this isn't hugely unexpected. If you send me your fio job files, I'll try and see what I can do about it.

Quoted 23 lines

Bisect down to patch:
1d2235152dc745c6d94bedb550fea84cffdbf768 is first bad commit commit 1d2235152dc745c6d94bedb550fea84cffdbf768

Author: Jens Axboe <[email protected]>
Date:   Fri Oct 2 19:27:04 2009 +0200

cfq-iosched: add a knob for desktop interactiveness

After I revert the patch against 2.6.23-rc3, fio_mmap_rand_read_4k regression disappears.

fio_mmap_randrw_4k has less than 20% regression and reverting the patch could restore performance.

On another stoakley machine with another JBOD, we see the similiar regression.

ffsb rand read/write has a bigger regression and reverting the patch could restore performance.

Yanmin

Signature · 1 line
Jens Axboe
#2 lore raw
Quoted 9 lines · Jens Axboe
Quoted 8 lines · Yanmin

Comparing with 2.6.23-rc1's result, fio rand read write has regression on my 2*4 core stoakley machine (8GB memory) with a JBOD of 12 disks.

Every disk has 8 1-GB files. Start 8 sub-processes per disk and every process random chooses a file on the disk to do 36 times of file read or write on the file and then choose another file.

Quoted 7 lines
Quoted 1 line

fio_mmap_rand_read_4k regresion is about 35%.

Heh, I seem to recollect I told Linus that this would cost is 30-40% performance. So not totally crazy.

So yes, this isn't hugely unexpected. If you send me your fio job files, I'll try and see what I can do about it.

Oh, and can you check whether setting

/sys/block/sdX/queue/iosched/low_latency

to 0 for the involved devices makes the regression go away?

Signature · 1 line
Jens Axboe
#3 lore raw
Quoted 15 lines · Jens Axboe
Quoted 14 lines · Jens Axboe
Quoted 9 lines · Yanmin

Comparing with 2.6.23-rc1's result, fio rand read write has regression on my 2*4 core stoakley machine (8GB memory) with a JBOD of 12 disks.

Every disk has 8 1-GB files. Start 8 sub-processes per disk and every process random chooses a file on the disk to do 36 times of file read or write on the file and then choose another file.

fio_mmap_rand_read_4k regresion is about 35%.

Heh, I seem to recollect I told Linus that this would cost is 30-40% performance. So not totally crazy.

Quoted 8 lines
Quoted 2 lines

So yes, this isn't hugely unexpected. If you send me your fio job files, I'll try and see what I can do about it.

Oh, and can you check whether setting

/sys/block/sdX/queue/iosched/low_latency

to 0 for the involved devices makes the regression go away?

Echo 0 to low_latency makes the regression become less than 3%.

#4 lore raw
In reply to Jens Axboe
Quoted 9 lines · Jens Axboe
Quoted 8 lines · Yanmin

Comparing with 2.6.23-rc1's result, fio rand read write has regression on my 2*4 core stoakley machine (8GB memory) with a JBOD of 12 disks.

Every disk has 8 1-GB files. Start 8 sub-processes per disk and every process random chooses a file on the disk to do 36 times of file read or write on the file and then choose another file.

Quoted 7 lines
Quoted 1 line

fio_mmap_rand_read_4k regresion is about 35%.

Heh, I seem to recollect I told Linus that this would cost is 30-40% performance. So not totally crazy.

So yes, this isn't hugely unexpected. If you send me your fio job files, I'll try and see what I can do about it.

See the attachment.

Quoted 25 lines
Quoted 24 lines

Bisect down to patch:
1d2235152dc745c6d94bedb550fea84cffdbf768 is first bad commit commit 1d2235152dc745c6d94bedb550fea84cffdbf768

Author: Jens Axboe <[email protected]>
Date:   Fri Oct 2 19:27:04 2009 +0200

cfq-iosched: add a knob for desktop interactiveness

After I revert the patch against 2.6.23-rc3, fio_mmap_rand_read_4k regression disappears.

fio_mmap_randrw_4k has less than 20% regression and reverting the patch could restore performance.

On another stoakley machine with another JBOD, we see the similiar regression.

ffsb rand read/write has a bigger regression and reverting the patch could restore performance.

Yanmin

#5 lore raw
Quoted 10 lines · Yanmin
Quoted 9 lines · Jens Axboe
Quoted 8 lines · Yanmin

Comparing with 2.6.23-rc1's result, fio rand read write has regression on my 2*4 core stoakley machine (8GB memory) with a JBOD of 12 disks.

Every disk has 8 1-GB files. Start 8 sub-processes per disk and every process random chooses a file on the disk to do 36 times of file read or write on the file and then choose another file.

Quoted 9 lines
Quoted 7 lines
Quoted 1 line

fio_mmap_rand_read_4k regresion is about 35%.

Heh, I seem to recollect I told Linus that this would cost is 30-40% performance. So not totally crazy.

So yes, this isn't hugely unexpected. If you send me your fio job files, I'll try and see what I can do about it.

See the attachment.

Hi Yanmin,
the fio test you sent just performs random read, no write seems involved here. I suspect that you should be able to observe the same regression if you just run on a single disk. Can you confirm?
Is your disk a SATA2 rotational disk with NCQ?

Thanks,
Corrado

Signature · 5 lines
__________________________________________________________________________

dott. Corrado Zoccolo                          mailto:[email protected]
PhD - Department of Computer Science - University of Pisa, Italy
--------------------------------------------------------------------------
#6 lore raw
Quoted 17 lines · Corrado Zoccolo
Quoted 15 lines · Yanmin
Quoted 14 lines · Jens Axboe
Quoted 9 lines · Yanmin

Comparing with 2.6.23-rc1's result, fio rand read write has regression on my 2*4 core stoakley machine (8GB memory) with a JBOD of 12 disks.

Every disk has 8 1-GB files. Start 8 sub-processes per disk and every process random chooses a file on the disk to do 36 times of file read or write on the file and then choose another file.

fio_mmap_rand_read_4k regresion is about 35%.

Heh, I seem to recollect I told Linus that this would cost is 30-40% performance. So not totally crazy.

Quoted 6 lines
Quoted 4 lines
Quoted 2 lines

So yes, this isn't hugely unexpected. If you send me your fio job files, I'll try and see what I can do about it.

See the attachment.

Hi Yanmin,
the fio test you sent just performs random read, no write seems involved here.


Sorry for not sending all the job files because they are similiar and big. You could change the job file to run rand write and readwrite.

Quoted 3 lines

I suspect that you should be able to observe the same regression if you just run on a single disk. Can you confirm?
Is your disk a SATA2 rotational disk with NCQ?

It's a JOBD with 12 SAS disks.

Another machine's JBOD has 13 SCSI disks.

Based on your suggestion, I tested it on a single SAS disk (create 8 1-GB file and start 8 processes) and get 30% regression. On a single SCSI disk, there is also about 30% regression. On a single SATA (SATA link up 1.5 Gbps) disk with NCQ, the regression is about 20%.

Perhaps the job file causes fio to deliver too many I/O requests, sometimes kernel print out some blocking info like below:

Apr  8 23:54:19 lkp-tulsa01-x8664 kernel: INFO: task kjournald:6261 blocked for more than 120 seconds.
Apr  8 23:54:19 lkp-tulsa01-x8664 kernel: "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
Apr  8 23:54:19 lkp-tulsa01-x8664 kernel: kjournald     D ffff88002f228180     0  6261      2
Apr  8 23:54:19 lkp-tulsa01-x8664 kernel:  ffff88027ab09910 0000000000000046 0000000000000000 ffff88027cc20440
Apr  8 23:54:19 lkp-tulsa01-x8664 kernel:  ffff88027f0c66a0 ffff88027ab09bb0 0000000100000e56 000000007cc20440
Apr  8 23:54:19 lkp-tulsa01-x8664 kernel:  ffff88027a915480 ffffffff803678a9 0000000000000000 0000000100000000
Apr  8 23:54:19 lkp-tulsa01-x8664 kernel: Call Trace:
Apr  8 23:54:19 lkp-tulsa01-x8664 kernel:  [<ffffffff803678a9>] ? __make_request+0x310/0x40b
Apr  8 23:54:19 lkp-tulsa01-x8664 kernel:  [<ffffffff802b5571>] ? sync_buffer+0x0/0x3f
Apr  8 23:54:19 lkp-tulsa01-x8664 kernel:  [<ffffffff805aeadf>] ? schedule+0x9/0x1d