[ltt-dev] [RFC PATCH] block: Fix bio merge induced high I/O latency

[Jens Axboe]

On Tue, Jan 20 2009, Jens Axboe wrote:
> On Mon, Jan 19 2009, Mathieu Desnoyers wrote:
> > * Jens Axboe (jens.axboe at oracle.com) wrote:
> > > On Sun, Jan 18 2009, Mathieu Desnoyers wrote:
> > > > I looked at the "ls" behavior (while doing a dd) within my LTTng trace
> > > > to create a fio job file.  The said behavior is appended below as "Part
> > > > 1 - ls I/O behavior". Note that the original "ls" test case was done
> > > > with the anticipatory I/O scheduler, which was active by default on my
> > > > debian system with custom vanilla 2.6.28 kernel. Also note that I am
> > > > running this on a raid-1, but have experienced the same problem on a
> > > > standard partition I created on the same machine.
> > > > 
> > > > I created the fio job file appended as "Part 2 - dd+ls fio job file". It
> > > > consists of one dd-like job and many small jobs reading as many data as
> > > > ls did. I used the small test script to batch run this ("Part 3 - batch
> > > > test").
> > > > 
> > > > The results for the ls-like jobs are interesting :
> > > > 
> > > > I/O scheduler        runt-min (msec)   runt-max (msec)
> > > > noop                       41             10563
> > > > anticipatory               63              8185
> > > > deadline                   52             33387
> > > > cfq                        43              1420
> > > 
> > 
> > Extra note : I have a HZ=250 on my system. Changing to 100 or 1000 did
> > not make much difference (also tried with NO_HZ enabled).
> > 
> > > Do you have queuing enabled on your drives? You can check that in
> > > /sys/block/sdX/device/queue_depth. Try setting those to 1 and retest all
> > > schedulers, would be good for comparison.
> > > 
> > 
> > Here are the tests with a queue_depth of 1 :
> > 
> > I/O scheduler        runt-min (msec)   runt-max (msec)
> > noop                       43             38235
> > anticipatory               44              8728
> > deadline                   51             19751
> > cfq                        48               427
> > 
> > 
> > Overall, I wouldn't say it makes much difference.
> 
> 0,5 seconds vs 1,5 seconds isn't much of a difference?
> 
> > > raid personalities or dm complicates matters, since it introduces a
> > > disconnect between 'ls' and the io scheduler at the bottom...
> > > 
> > 
> > Yes, ideally I should re-run those directly on the disk partitions.
> 
> At least for comparison.
> 
> > I am also tempted to create a fio job file which acts like a ssh server
> > receiving a connexion after it has been pruned from the cache while the
> > system if doing heavy I/O. "ssh", in this case, seems to be doing much
> > more I/O than a simple "ls", and I think we might want to see if cfq
> > behaves correctly in such case. Most of this I/O is coming from page
> > faults (identified as traps in the trace) probably because the ssh
> > executable has been thrown out of the cache by
> > 
> > echo 3 > /proc/sys/vm/drop_caches
> > 
> > The behavior of an incoming ssh connexion after clearing the cache is
> > appended below (Part 1 - LTTng trace for incoming ssh connexion). The
> > job file created (Part 2) reads, for each job, a 2MB file with random
> > reads each between 4k-44k. The results are very interesting for cfq :
> > 
> > I/O scheduler        runt-min (msec)   runt-max (msec)
> > noop                       586           110242
> > anticipatory               531            26942
> > deadline                   561           108772
> > cfq                        523            28216
> > 
> > So, basically, ssh being out of the cache can take 28s to answer an
> > incoming ssh connexion even with the cfq scheduler. This is not exactly
> > what I would call an acceptable latency.
> 
> At some point, you have to stop and consider what is acceptable
> performance for a given IO pattern. Your ssh test case is purely random
> IO, and neither CFQ nor AS would do any idling for that. We can make
> this test case faster for sure, the hard part is making sure that we
> don't regress on async throughput at the same time.
> 
> Also remember that with your raid1, it's not entirely reasonable to
> blaim all performance issues on the IO scheduler as per my previous
> mail. It would be a lot more fair to view the disk numbers individually.
> 
> Can you retry this job with 'quantum' set to 1 and 'slice_async_rq' set
> to 1 as well?
> 
> However, I think we should be doing somewhat better at this test case.

Mathieu, does this improve anything for you?

diff --git a/block/cfq-iosched.c b/block/cfq-iosched.c
index e8525fa..a556512 100644
--- a/block/cfq-iosched.c
+++ b/block/cfq-iosched.c
@@ -1765,6 +1765,32 @@ cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 }
 
 /*
+ * Pull dispatched requests from 'cfqq' back into the scheduler
+ */
+static void cfq_pull_dispatched_requests(struct cfq_data *cfqd,
+					 struct cfq_queue *cfqq)
+{
+	struct request_queue *q = cfqd->queue;
+	struct request *rq, *tmp;
+
+	list_for_each_entry_safe(rq, tmp, &q->queue_head, queuelist) {
+		if ((rq->cmd_flags & REQ_STARTED) || RQ_CFQQ(rq) != cfqq)
+			continue;
+
+		/*
+		 * Pull off the dispatch list and put it back into the cfqq
+		 */
+		list_del(&rq->queuelist);
+		cfqq->dispatched--;
+		if (cfq_cfqq_sync(cfqq))
+			cfqd->sync_flight--;
+
+		list_add_tail(&rq->queuelist, &cfqq->fifo);
+		cfq_add_rq_rb(rq);
+	}
+}
+
+/*
  * Check if new_cfqq should preempt the currently active queue. Return 0 for
  * no or if we aren't sure, a 1 will cause a preempt.
  */
@@ -1820,8 +1846,14 @@ cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
  */
 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
+	struct cfq_queue *old_cfqq = cfqd->active_queue;
+
 	cfq_log_cfqq(cfqd, cfqq, "preempt");
-	cfq_slice_expired(cfqd, 1);
+
+	if (old_cfqq) {
+		__cfq_slice_expired(cfqd, old_cfqq, 1);
+		cfq_pull_dispatched_requests(cfqd, old_cfqq);
+	}
 
 	/*
 	 * Put the new queue at the front of the of the current list,

-- 
Jens Axboe