public inbox for [email protected]
 help / color / mirror / Atom feed
From: Pavel Begunkov <[email protected]>
To: Hao Xu <[email protected]>, Jens Axboe <[email protected]>
Cc: [email protected], Joseph Qi <[email protected]>
Subject: Re: [PATCH v6 0/6] task work optimization
Date: Fri, 3 Dec 2021 01:39:55 +0000	[thread overview]
Message-ID: <[email protected]> (raw)
In-Reply-To: <[email protected]>

On 11/26/21 10:07, Hao Xu wrote:
> v4->v5
> - change the implementation of merge_wq_list
> 
> v5->v6
> - change the logic of handling prior task list to:
>    1) grabbed uring_lock: leverage the inline completion infra
>    2) otherwise: batch __req_complete_post() calls to save
>       completion_lock operations.

some testing for v6, first is taking first 5 patches (1-5), and
then all 6 (see 1-6).

modprobe null_blk no_sched=1 irqmode=1 completion_nsec=0 submit_queues=16 poll_queues=32 hw_queue_depth=128
echo 2 | sudo tee /sys/block/nullb0/queue/nomerges
echo 0 | sudo tee /sys/block/nullb0/queue/iostats
mitigations=off

added this to test non-sqpoll:

@@ -2840,7 +2840,7 @@ static void io_complete_rw(struct kiocb *kiocb, long res)
                 return;
         req->result = res;
         req->io_task_work.func = io_req_task_complete;
-       io_req_task_work_add(req, !!(req->ctx->flags & IORING_SETUP_SQPOLL));
+       io_req_task_work_add(req, true);
  }

# 1-5, sqpoll=0
nice -n -20 taskset -c 0 ./io_uring -d32 -s32 -c32 -p0 -B1 -F1 -b512 /dev/nullb0
IOPS=3238688, IOS/call=32/32, inflight=32 (32)
IOPS=3299776, IOS/call=32/32, inflight=32 (32)
IOPS=3328416, IOS/call=32/32, inflight=32 (32)
IOPS=3291488, IOS/call=32/32, inflight=32 (32)
IOPS=3284480, IOS/call=32/32, inflight=32 (32)
IOPS=3305248, IOS/call=32/32, inflight=32 (32)
IOPS=3275392, IOS/call=32/32, inflight=32 (32)
IOPS=3301376, IOS/call=32/32, inflight=32 (32)
IOPS=3287392, IOS/call=32/32, inflight=32 (32)

# 1-5, sqpoll=1
nice -n -20  ./io_uring -d32 -s32 -c32 -p0 -B1 -F1 -b512 /dev/nullb0
IOPS=2730752, IOS/call=2730752/2730752, inflight=32 (32)
IOPS=2822432, IOS/call=-1/-1, inflight=0 (32)
IOPS=2818464, IOS/call=-1/-1, inflight=32 (32)
IOPS=2802880, IOS/call=-1/-1, inflight=0 (32)
IOPS=2773440, IOS/call=-1/-1, inflight=32 (32)
IOPS=2827296, IOS/call=-1/-1, inflight=32 (32)
IOPS=2808320, IOS/call=-1/-1, inflight=32 (32)
IOPS=2793120, IOS/call=-1/-1, inflight=32 (32)
IOPS=2769632, IOS/call=-1/-1, inflight=32 (32)
IOPS=2752896, IOS/call=-1/-1, inflight=32 (32)

# 1-6, sqpoll=0
nice -n -20 taskset -c 0 ./io_uring -d32 -s32 -c32 -p0 -B1 -F1 -b512 /dev/nullb0
IOPS=3219552, IOS/call=32/32, inflight=32 (32)
IOPS=3284128, IOS/call=32/32, inflight=32 (32)
IOPS=3305024, IOS/call=32/32, inflight=32 (32)
IOPS=3301920, IOS/call=32/32, inflight=32 (32)
IOPS=3330592, IOS/call=32/32, inflight=32 (32)
IOPS=3286496, IOS/call=32/32, inflight=32 (32)
IOPS=3236160, IOS/call=32/32, inflight=32 (32)
IOPS=3307552, IOS/call=32/32, inflight=32 (32)

# 1-6, sqpoll=1
nice -n -20  ./io_uring -d32 -s32 -c32 -p0 -B1 -F1 -b512 /dev/nullb0
IOPS=2777152, IOS/call=2777152/2777152, inflight=32 (32)
IOPS=2822080, IOS/call=-1/-1, inflight=32 (32)
IOPS=2785472, IOS/call=-1/-1, inflight=0 (32)
IOPS=2763360, IOS/call=-1/-1, inflight=0 (32)
IOPS=2789856, IOS/call=-1/-1, inflight=32 (32)
IOPS=2783296, IOS/call=-1/-1, inflight=32 (32)
IOPS=2786016, IOS/call=-1/-1, inflight=0 (32)
IOPS=2773760, IOS/call=-1/-1, inflight=32 (32)
IOPS=2745408, IOS/call=-1/-1, inflight=32 (32)
IOPS=2764352, IOS/call=-1/-1, inflight=32 (32)
IOPS=2766912, IOS/call=-1/-1, inflight=32 (32)
IOPS=2757216, IOS/call=-1/-1, inflight=32 (32)

So, no difference here as expected, it just takes uring_lock
as per v6 changes and goes through the old path. Than I added
this to compare old vs new paths:

@@ -2283,7 +2283,7 @@ static void handle_prior_tw_list(struct io_wq_work_node *node, struct io_ring_ct
                         ctx_flush_and_put(*ctx, locked);
                         *ctx = req->ctx;
                         /* if not contended, grab and improve batching */
-                       *locked = mutex_trylock(&(*ctx)->uring_lock);
+                       // *locked = mutex_trylock(&(*ctx)->uring_lock);
                         percpu_ref_get(&(*ctx)->refs);
                         if (unlikely(!*locked))
                                 spin_lock(&(*ctx)->completion_lock);


# 1-6 + no trylock, sqpoll=0
nice -n -20 taskset -c 0 ./io_uring -d32 -s32 -c32 -p0 -B1 -F1 -b512 /dev/nullb0
IOPS=3239040, IOS/call=32/32, inflight=32 (32)
IOPS=3244800, IOS/call=32/32, inflight=32 (32)
IOPS=3208544, IOS/call=32/32, inflight=32 (32)
IOPS=3264384, IOS/call=32/32, inflight=32 (32)
IOPS=3264000, IOS/call=32/32, inflight=32 (32)
IOPS=3296960, IOS/call=32/32, inflight=32 (32)
IOPS=3283424, IOS/call=32/32, inflight=32 (32)
IOPS=3284064, IOS/call=32/32, inflight=32 (32)
IOPS=3275232, IOS/call=32/32, inflight=32 (32)
IOPS=3261248, IOS/call=32/32, inflight=32 (32)
IOPS=3273792, IOS/call=32/32, inflight=32 (32)

#1-6 + no trylock, sqpoll=1
nice -n -20  ./io_uring -d32 -s32 -c32 -p0 -B1 -F1 -b512 /dev/nullb0
IOPS=2676736, IOS/call=2676736/2676736, inflight=32 (32)
IOPS=2639776, IOS/call=-1/-1, inflight=32 (32)
IOPS=2660000, IOS/call=-1/-1, inflight=32 (32)
IOPS=2639584, IOS/call=-1/-1, inflight=32 (32)
IOPS=2634592, IOS/call=-1/-1, inflight=0 (32)
IOPS=2611488, IOS/call=-1/-1, inflight=32 (32)
IOPS=2647360, IOS/call=-1/-1, inflight=32 (32)
IOPS=2630720, IOS/call=-1/-1, inflight=32 (32)
IOPS=2663200, IOS/call=-1/-1, inflight=32 (32)
IOPS=2694240, IOS/call=-1/-1, inflight=32 (32)
IOPS=2674592, IOS/call=-1/-1, inflight=32 (32)

Seems it goes a little bit down, but not much. Considering that
it's an optimisation for cases where there is no batching at all,
that's good.

-- 
Pavel Begunkov

  parent reply	other threads:[~2021-12-03  1:39 UTC|newest]

Thread overview: 22+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-11-26 10:07 [PATCH v6 0/6] task work optimization Hao Xu
2021-11-26 10:07 ` [PATCH 1/6] io-wq: add helper to merge two wq_lists Hao Xu
2021-11-26 10:07 ` [PATCH 2/6] io_uring: add a priority tw list for irq completion work Hao Xu
2021-11-26 10:07 ` [PATCH 3/6] io_uring: add helper for task work execution code Hao Xu
2021-11-26 10:07 ` [PATCH 4/6] io_uring: split io_req_complete_post() and add a helper Hao Xu
2021-11-26 10:07 ` [PATCH 5/6] io_uring: move up io_put_kbuf() and io_put_rw_kbuf() Hao Xu
2021-11-26 10:07 ` [PATCH 6/6] io_uring: batch completion in prior_task_list Hao Xu
2021-11-26 12:56   ` Hao Xu
2021-11-26 13:37 ` [PATCH RESEND " Hao Xu
2021-11-27 15:24   ` [PATCH v7] " Hao Xu
2021-11-28 15:28     ` Pavel Begunkov
2021-12-03  1:39 ` Pavel Begunkov [this message]
2021-12-03  2:01   ` [PATCH v6 0/6] task work optimization Pavel Begunkov
2021-12-03  7:30     ` Hao Xu
2021-12-03 14:21       ` Pavel Begunkov
2021-12-05 15:02         ` Hao Xu
2021-12-05 15:42           ` Pavel Begunkov
2021-12-06  8:35             ` Hao Xu
2021-12-06  9:48               ` Hao Xu
2021-12-03  3:24   ` Hao Xu
2021-12-04 20:58 ` Pavel Begunkov
2021-12-05 15:11   ` Hao Xu

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    [email protected] \
    [email protected] \
    [email protected] \
    [email protected] \
    [email protected] \
    [email protected] \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox