Re: [PATCH RFC - TAKE TWO - 10/12] block, bfq: add Early Queue Merge (EQM)

[Paolo Valente]

Il giorno 03/giu/2014, alle ore 18:28, Tejun Heo <tj@xxxxxxxxxx> ha scritto:

> Hello,
> 
> On Mon, Jun 02, 2014 at 11:46:45AM +0200, Paolo Valente wrote:
>>> I don't really follow the last part.  So, the difference is that
>>> cooperating queue setup also takes place during bio merge too, right?
>> 
>> Not only, in bfq an actual queue merge is performed in the bio-merge hook.
> 
> I think I'm a bit confused because it's named "early" queue merge
> while it actually moves queue merging later than cfq - set_request()
> happens before bio/rq merging.


There is probably something I am missing here, because, as can be seen in blk-core.c,
around line 1495, elv_set_request() is invoked in the context of the get_request() function,
which in its turn is called from blk_queue_bio() *after* attempting both a plug merge
and a merge with one of the requests in the block layer's cache. The first
attempt is lockless and doesn't involve the I/O scheduler, but the
second attempt includes invoking the allow_merge_fn hook of the scheduler
(elv_merge() -> elv_rq_merge_ok() -> elv_iosched_allow_merge()).

Furthermore, as far as I know, it is true that CFQ actually merges queues in the
set_request hook, but a cooperator is searched for a queue (and, if it is found,
the two queues are scheduled to merge) only when the queue expires after being
served (see cfq_select_queue() and the two functions cfq_close_cooperator() and
cfq_setup_merge() that it invokes). If a cooperator is found, it is forcedly
served; however, the actual merge of the two queues happens at the next
set_request (cfq_merge_bfqqs()).

In contrast, BFQ both searches for a cooperator and merges the queue with a
newly-found cooperator in the allow_merge hook, which is "earlier" with respect
to CFQ, as it doesn't need to wait for a queue to be served and expire, and for its
associated process to issue new I/O. Hence the name Early Queue Merge.


> So, what it tries to do is
> compensating for the lack of cfq_rq_close() preemption at request
> issue time, right?
> 

Yes, thanks to early merging, there is then no need to recover a lost sequential
pattern through preemptions.

>>> cfq does it once when allocating the request.  That seems a lot more
>>> reasonable to me.  It's doing that once for one start sector.  I mean,
>>> plugging is usually extremely short compared to actual IO service
>>> time.  It's there to mask the latencies between bio issues that the
>>> same CPU is doing.  I can't see how this earliness can be actually
>>> useful.  Do you have results to back this one up?  Or is this just
>>> born out of thin air?
>> 
>> Arianna added the early-queue-merge part in the allow_merge_fn hook
>> about one year ago, as a a consequence of a throughput loss of about
>> 30% with KVM/QEMU workloads. In particular, we ran most of the tests
>> on a WDC WD60000HLHX-0 Velociraptor. That HDD might not be available
>> for testing any more, but we can reproduce our results for you on
>> other HDDs, with and without early queue merge. And, maybe through
>> traces, we can show you that the reason for the throughput loss is
>> exactly that described (in a wordy way) in this patch. Of course
>> unless we have missed something.
> 
> Oh, as long as it makes measureable difference, I have no objection;
> however, I do think more explanation and comments would be nice.  I
> still can't quite understand why retrying on each merge attempt would
> make so much difference.  Maybe I just failed to understand what you
> wrote in the commit message.

If we remember well, one of the problems was exactly that a different request
may become the head request of the in-service queue between two rq merge
attempts. If we do not retry on every attempt, we lose the chance
to merge the queue at hand with the in-service queue. The two queues may
then diverge, and hence have no other opportunity to be merged.

> Is it because the cooperating tasks
> issue IOs which grow large and close enough after merges but not on
> the first bio issuance?  If so, why isn't doing it on rq merge time
> enough?  Is the timing sensitive enough for certain workloads that
> waiting till unplug time misses the opportunity?  But plugging should
> be relatively short compared to the time actual IOs take, so why would
> it be that sensitive?  What am I missing here?

The problem is not the duration of the plugging, but the fact that, if a request merge
succeeds for a bio, then there will be no set_request invocation for that bio.
Therefore, without early merging, there will be no queue merge at all.

If my replies are correct and convince you, then I will use them to integrate and
hopefully improve the documentation for this patch.

Paolo

> 
> Thanks.
> 
> -- 
> tejun


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/