Skip to contentSkip to conversation

cfq-iosched: quantum check tweak --resend

PATCH In review Shaohua Li 2 participants 6 messages Mar 1, 2010 Block layer

lore
Share & tools
lore

Shaohua Li

author
#0 lore raw

PATCH 1/1 cfq-iosched: quantum check tweak --resend

Currently a queue can only dispatch up to 4 requests if there are other queues. This isn't optimal, device can handle more requests, for example, AHCI can handle 31 requests. I can understand the limit is for fairness, but we could do a tweak: if the queue still has a lot of slice left, sounds we could ignore the limit. Test shows this boost my workload (two thread randread of a SSD) from 78m/s to 100m/s.
Thanks for suggestions from Corrado and Vivek for the patch.

---
 block/cfq-iosched.c |   30 ++++++++++++++++++++++++++----
 1 files changed, 26 insertions(+), 4 deletions(-)
block/cfq-iosched.c +26 -4
index f27e535..0db07d7 100644
@@ -19,7 +19,7 @@
  * tunables
  */
 /* max queue in one round of service */
-static const int cfq_quantum = 4;
+static const int cfq_quantum = 8;
 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
 /* maximum backwards seek, in KiB */
 static const int cfq_back_max = 16 * 1024;
@@ -2197,6 +2197,19 @@ static int cfq_forced_dispatch(struct cfq_data *cfqd)
 	return dispatched;
 }
 
+static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
+	struct cfq_queue *cfqq)
+{
+	/* the queue hasn't finished any request, can't estimate */
+	if (cfq_cfqq_slice_new(cfqq))
+		return 1;
+	if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
+		cfqq->slice_end))
+		return 1;
+
+	return 0;
+}
+
 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
 	unsigned int max_dispatch;
@@ -2213,7 +2226,7 @@ static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 	if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
 		return false;
 
-	max_dispatch = cfqd->cfq_quantum;
+	max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
 	if (cfq_class_idle(cfqq))
 		max_dispatch = 1;
 
@@ -2230,13 +2243,22 @@ static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 		/*
 		 * We have other queues, don't allow more IO from this one
 		 */
-		if (cfqd->busy_queues > 1)
+		if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
 			return false;
 
 		/*
 		 * Sole queue user, no limit
 		 */
-		max_dispatch = -1;
+		if (cfqd->busy_queues == 1)
+			max_dispatch = -1;
+		else
+			/*
+			 * Normally we start throttling cfqq when cfq_quantum/2
+			 * requests have been dispatched. But we can drive
+			 * deeper queue depths at the beginning of slice
+			 * subjected to upper limit of cfq_quantum.
+			 * */
+			max_dispatch = cfqd->cfq_quantum;
 	}
 
 	/*
Signature · 1 line
1.6.3.3
#1 lore raw
Quoted 7 lines · Shaohua Li

Currently a queue can only dispatch up to 4 requests if there are other queues. This isn't optimal, device can handle more requests, for example, AHCI can handle 31 requests. I can understand the limit is for fairness, but we could do a tweak: if the queue still has a lot of slice left, sounds we could ignore the limit. Test shows this boost my workload (two thread randread of a SSD) from 78m/s to 100m/s.
Thanks for suggestions from Corrado and Vivek for the patch.

As mentioned before, I think we definitely want to ensure that we drive the full queue depth whenever possible. I think your patch is a bit dangerous, though. The problematic workload here is a buffered write, interleaved with the occasional sync reader. If the sync reader has to endure 32 requests every time, latency rises dramatically for him.

Signature · 1 line
Jens Axboe
author
#2 lore raw
Quoted 9 lines · Jens Axboe
Quoted 7 lines · Shaohua Li

Currently a queue can only dispatch up to 4 requests if there are other queues. This isn't optimal, device can handle more requests, for example, AHCI can handle 31 requests. I can understand the limit is for fairness, but we could do a tweak: if the queue still has a lot of slice left, sounds we could ignore the limit. Test shows this boost my workload (two thread randread of a SSD) from 78m/s to 100m/s.
Thanks for suggestions from Corrado and Vivek for the patch.

Quoted 5 lines

As mentioned before, I think we definitely want to ensure that we drive the full queue depth whenever possible. I think your patch is a bit dangerous, though. The problematic workload here is a buffered write, interleaved with the occasional sync reader. If the sync reader has to endure 32 requests every time, latency rises dramatically for him.

the patch still matains a hardlimit for dispatched request. For a async, the limit is cfq_slice_async/cfq_slice_idle = 5. For sync, the limit is 8. And we only pipe out such number of requests at the begining of a slice. For the workload you mentioned here, we only dispatch 1 extra request.

Thanks,
Shaohua

#3 lore raw
Quoted 15 lines · Shaohua Li
Quoted 9 lines · Jens Axboe
Quoted 7 lines · Shaohua Li

Currently a queue can only dispatch up to 4 requests if there are other queues. This isn't optimal, device can handle more requests, for example, AHCI can handle 31 requests. I can understand the limit is for fairness, but we could do a tweak: if the queue still has a lot of slice left, sounds we could ignore the limit. Test shows this boost my workload (two thread randread of a SSD) from 78m/s to 100m/s.
Thanks for suggestions from Corrado and Vivek for the patch.

Quoted 5 lines

As mentioned before, I think we definitely want to ensure that we drive the full queue depth whenever possible. I think your patch is a bit dangerous, though. The problematic workload here is a buffered write, interleaved with the occasional sync reader. If the sync reader has to endure 32 requests every time, latency rises dramatically for him.

Quoted 4 lines

the patch still matains a hardlimit for dispatched request. For a async, the limit is cfq_slice_async/cfq_slice_idle = 5. For sync, the limit is 8. And we only pipe out such number of requests at the begining of a slice. For the workload you mentioned here, we only dispatch 1 extra request.

OK, that sound appropriate. Final question - why change the quantum and use quantum/2?

Signature · 1 line
Jens Axboe
author
#4 lore raw
Quoted 16 lines · Jens Axboe
Quoted 15 lines · Shaohua Li
Quoted 14 lines · Jens Axboe
Quoted 7 lines · Shaohua Li

Currently a queue can only dispatch up to 4 requests if there are other queues. This isn't optimal, device can handle more requests, for example, AHCI can handle 31 requests. I can understand the limit is for fairness, but we could do a tweak: if the queue still has a lot of slice left, sounds we could ignore the limit. Test shows this boost my workload (two thread randread of a SSD) from 78m/s to 100m/s.
Thanks for suggestions from Corrado and Vivek for the patch.

As mentioned before, I think we definitely want to ensure that we drive the full queue depth whenever possible. I think your patch is a bit dangerous, though. The problematic workload here is a buffered write, interleaved with the occasional sync reader. If the sync reader has to endure 32 requests every time, latency rises dramatically for him.

Quoted 7 lines
Quoted 4 lines

the patch still matains a hardlimit for dispatched request. For a async, the limit is cfq_slice_async/cfq_slice_idle = 5. For sync, the limit is 8. And we only pipe out such number of requests at the begining of a slice. For the workload you mentioned here, we only dispatch 1 extra request.

OK, that sound appropriate. Final question - why change the quantum and use quantum/2?

This is suggested by Vivek. In this way quantum is still the hard limit and doesn't surprise users. we do throttling at 1/2 quantum (softlimit) and then stop at quantum (hard limit)

Thanks,
Shaohua

#5 lore raw
Quoted 22 lines · Shaohua Li
Quoted 21 lines · Jens Axboe
Quoted 19 lines · Shaohua Li
Quoted 9 lines · Jens Axboe
Quoted 7 lines · Shaohua Li

Currently a queue can only dispatch up to 4 requests if there are other queues. This isn't optimal, device can handle more requests, for example, AHCI can handle 31 requests. I can understand the limit is for fairness, but we could do a tweak: if the queue still has a lot of slice left, sounds we could ignore the limit. Test shows this boost my workload (two thread randread of a SSD) from 78m/s to 100m/s.
Thanks for suggestions from Corrado and Vivek for the patch.

Quoted 5 lines

As mentioned before, I think we definitely want to ensure that we drive the full queue depth whenever possible. I think your patch is a bit dangerous, though. The problematic workload here is a buffered write, interleaved with the occasional sync reader. If the sync reader has to endure 32 requests every time, latency rises dramatically for him.

the patch still matains a hardlimit for dispatched request. For a async, the limit is cfq_slice_async/cfq_slice_idle = 5. For sync, the limit is 8. And we only pipe out such number of requests at the begining of a slice. For the workload you mentioned here, we only dispatch 1 extra request.

Quoted 5 lines
Quoted 2 lines

OK, that sound appropriate. Final question - why change the quantum and use quantum/2?

This is suggested by Vivek. In this way quantum is still the hard limit and doesn't surprise users. we do throttling at 1/2 quantum (softlimit) and then stop at quantum (hard limit)

OK, that makes sense. I will apply the patch, thanks!

Signature · 1 line
Jens Axboe