Re: [PATCH] blk-mq: plug request for shared sbitmap

On 18/05/2021 12:16, Ming Lei wrote:

On Tue, May 18, 2021 at 10:44:43AM +0100, John Garry wrote:

quoted

On 14/05/2021 03:20, Ming Lei wrote:

quoted

In case of shared sbitmap, request won't be held in plug list any more
sine commit 32bc15afed04 ("blk-mq: Facilitate a shared sbitmap per
tagset"), this way makes request merge from flush plug list & batching
submission not possible, so cause performance regression.

Yanhui reports performance regression when running sequential IO
test(libaio, 16 jobs, 8 depth for each job) in VM, and the VM disk
is emulated with image stored on xfs/megaraid_sas.

Fix the issue by recovering original behavior to allow to hold request
in plug list.

Hi Ming,

Since testing v5.13-rc2, I noticed that this patch made the hang I was
seeing disappear:
https://lore.kernel.org/linux-scsi/3d72d64d-314f-9d34-e039-7e508b2abe1b@huawei.com/ (local)

I don't think that problem is solved, though.

This kind of hang or lockup is usually related with cpu utilization, and
this patch may reduce cpu utilization in submission context.

Right.

quoted

So I wonder about throughput performance (I had hoped to test before it was
merged). I only have 6x SAS SSDs at hand, but I see some significant changes
(good and bad) for mq-deadline for hisi_sas:
Before 620K (read), 300K IOPs (randread)
After 430K (read), 460-490K IOPs (randread)

'Before 620K' could be caused by busy queue when batching submission isn't
applied, so merge chance is increased. This patch applies batching
submission, so queue becomes not busy enough.

BTW, what is the queue depth of sdev and can_queue of shost for your hisilision SAS?

sdev queue depth is 64 (see hisi_sas_slave_configure()) and host depth 
is 4096 - 96 (for reserved tags) = 4000

IIRC, megaraid sas and mpt3sas use 256 for SAS sdev queue depth

quoted

none IO sched is always about 450K (read) and 500K (randread)

Do you guys have any figures? Are my results as expected?

In yanhui's virt workload(qemu, libaio, dio, high queue depth, single
job), the patch can improve throughput much(>50%) when running
sequential write(dio, libaio, 16 jobs) to XFS. And it is observed that
IO merge is recovered to level of disabling host tags.

Thanks,
John