Re: [PATCH 2/2] blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues

[Jan Kara <jack@suse.cz>]

On Tue 22-12-20 11:18:22, Jan Kara wrote:
> On Sat 19-12-20 11:14:27, Ming Lei wrote:
> > On Fri, Dec 18, 2020 at 10:44:12PM +0100, Jan Kara wrote:
> > > Currently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for
> > > a queue with multiple HW queues, the performance it rather bad. The
> > > problem is that these IO schedulers use queue-wide locking and their
> > > dispatch function does not respect the hctx it is passed in and returns
> > > any request it finds appropriate. Thus locality of request access is
> > > broken and dispatch from multiple CPUs just contends on IO scheduler
> > > locks. For these IO schedulers there's little point in dispatching from
> > > multiple CPUs. Instead dispatch always only from a single CPU to limit
> > > contention.
> > > 
> > > Below is a comparison of dbench runs on XFS filesystem where the storage
> > > is a raid card with 64 HW queues and to it attached a single rotating
> > > disk. BFQ is used as IO scheduler:
> > > 
> > >       clients           MQ                     SQ             MQ-Patched
> > > Amean 1      39.12 (0.00%)       43.29 * -10.67%*       36.09 *   7.74%*
> > > Amean 2     128.58 (0.00%)      101.30 *  21.22%*       96.14 *  25.23%*
> > > Amean 4     577.42 (0.00%)      494.47 *  14.37%*      508.49 *  11.94%*
> > > Amean 8     610.95 (0.00%)      363.86 *  40.44%*      362.12 *  40.73%*
> > > Amean 16    391.78 (0.00%)      261.49 *  33.25%*      282.94 *  27.78%*
> > > Amean 32    324.64 (0.00%)      267.71 *  17.54%*      233.00 *  28.23%*
> > > Amean 64    295.04 (0.00%)      253.02 *  14.24%*      242.37 *  17.85%*
> > > Amean 512 10281.61 (0.00%)    10211.16 *   0.69%*    10447.53 *  -1.61%*
> > > 
> > > Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is
> > > the same kernel with megaraid_sas.host_tagset_enable=0 so that the card
> > > advertises just a single HW queue. MQ-Patched is a kernel with this
> > > patch applied.
> > > 
> > > You can see multiple hardware queues heavily hurt performance in
> > > combination with BFQ. The patch restores the performance.
> > > 
> > > Signed-off-by: Jan Kara <jack@suse.cz>
> > > ---
> > >  block/blk-mq.c           | 37 +++++++++++++++++++++++++++++++++++++
> > >  block/kyber-iosched.c    |  1 +
> > >  include/linux/elevator.h |  2 ++
> > >  3 files changed, 40 insertions(+)
> > > 
> > > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > > index 57d0461f2be5..6d80054c231b 100644
> > > --- a/block/blk-mq.c
> > > +++ b/block/blk-mq.c
> > > @@ -1663,6 +1663,31 @@ void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
> > >  }
> > >  EXPORT_SYMBOL(blk_mq_run_hw_queue);
> > >  
> > > +static struct blk_mq_hw_ctx *blk_mq_sq_iosched_hctx(struct request_queue *q)
> > > +{
> > > +	struct elevator_queue *e = q->elevator;
> > > +	struct blk_mq_hw_ctx *hctx;
> > > +
> > > +	/*
> > > +	 * The queue has multiple hardware queues but uses IO scheduler that
> > > +	 * does not respect hardware queues when dispatching? This is not a
> > > +	 * great setup but it can be sensible when we have a single rotational
> > > +	 * disk behind a raid card. Just don't bother with multiple HW queues
> > > +	 * and dispatch from hctx for the current CPU since running multiple
> > > +	 * queues just causes lock contention inside the scheduler and
> > > +	 * pointless cache bouncing because the hctx is not respected by the IO
> > > +	 * scheduler's dispatch function anyway.
> > > +	 */
> > > +	if (q->nr_hw_queues > 1 && e && e->type->ops.dispatch_request &&
> > > +	    !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE)) {
> > > +		hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT,
> > > +					     raw_smp_processor_id());
> > > +		if (!blk_mq_hctx_stopped(hctx))
> > > +			return hctx;
> > > +	}
> > > +	return NULL;
> > > +}
> > > +
> > >  /**
> > >   * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
> > >   * @q: Pointer to the request queue to run.
> > > @@ -1673,6 +1698,12 @@ void blk_mq_run_hw_queues(struct request_queue *q, bool async)
> > >  	struct blk_mq_hw_ctx *hctx;
> > >  	int i;
> > >  
> > > +	hctx = blk_mq_sq_iosched_hctx(q);
> > > +	if (hctx) {
> > > +		blk_mq_run_hw_queue(hctx, async);
> > > +		return;
> > > +	}
> > > +
> > 
> > This approach looks reasonable, just wondering which code path is wrt.
> > blk_mq_run_hw_queues() improvement by this patch.
> > 
> > Since ed5dd6a67d5e ("scsi: core: Only re-run queue in scsi_end_request() if device
> > queue is busy") is merged, blk_mq_run_hw_queues() is only called from scsi_end_request()
> > when the scsi device is busy for megaraid.
> > 
> > Another one is bfq_schedule_dispatch(), in which blk_mq_run_hw_queues()
> > is still be called, if that is the reason, maybe it is easier to optimize
> > bfq_schedule_dispatch() by avoiding to call blk_mq_run_hw_queues().
> 
> That's a good question. Tracing shows that with dbench I'm seeing *lots*
> (about 23000/s) blk_mq_delay_run_hw_queues() calls, mostly from
> __blk_mq_do_dispatch_sched(). This drops to "only" about 2000 calls/s with
> my patches applied.
> 
> So it means BFQ decided not to dispatch any request (e.g. because it is
> idling for more IO from the same process) and that triggers that path in
> __blk_mq_do_dispatch_sched() that just queues the dispatch again. So blk-mq
> ends up polling BFQ rather heavily for requests it doesn't want to give out
> :). In this sense my patch just makes the real problem less severe.
> 
> I've noticed that if ->has_work() returned false, we would not end up
> calling blk_mq_delay_run_hw_queues(). But for BFQ ->has_work() often
> returns true because it has requests queued but ->dispatch_request()
> doesn't dispatch anything because of other scheduling constraints. And so
> we end up calling blk_mq_delay_run_hw_queues() because if we allocated
> dispatch budget and didn't dispatch in the end, we could have blocked
> dispatch from another hctx and so now need to rerun that hctx to dispatch
> possibly queued requests.
> 
> I was thinking how we could possibly improve this. One obvious possibility
> is to modify IO schedulers so that their ->has_work() does not return true
> if they later decide not to dispatch anything. However this can happen both
> to mq-deadline and BFQ and for either of them determining whether they will
> dispatch a request or not is about as expensive as dispatching it. So it
> doesn't seem very appealing for these IO schedulers to do the work twice or
> to somehow cache the request found. What seems more workable would be for
> blk_mq_put_dispatch_budget() to return whether rerunning the queue might be
> needed or not (for SCSI, which is the only subsystem using budgeting, this
> means returning whether we were currently at queue_depth) and use that
> information in __blk_mq_do_dispatch_sched(). I'll experiment with a patch I
> guess...

OK, so I was experimenting more with this. I've implemented my idea of
->put_budget() returning whether it is needed to rerun the queue (and then
using this inside __blk_mq_do_dispatch_sched()). This indeed mostly got rid
of calls of blk_mq_delay_run_hw_queues() from __blk_mq_do_dispatch_sched().
This patch on its own however was not enough to fix the regression when
megaraid_sas started using multiple HW queues - it got back about half of
the regression. There was still enough contention and cache bouncing from
the remaining blk_mq_run_hw_queues() calls - mostly called when a request was
completed and there was no request currently running (this happens in
bfq_completed_request -> bfq_schedule_dispatch path). And these calls seem
to be really needed AFAICT. Also I did test run with both the previous
series and "budget" patch applied and it didn't show any significant
difference to just the previous series. So although conceptually the patch
makes sense and reduces calls to blk_mq_delay_run_hw_queues(), I'm not sure
it is worth it given it brings no measurable benefit. What do people think?

								Honza

PS: I've noticed that my original patch is slightly buggy and we probably
need to run not only the hctx on current CPU but also any hctx with
non-empty ->dispatch list. Otherwise we could stall some requests. I'll fix
that up on next posting.

-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR