On Sat, May 09, 2020 at 04:57:48PM +0800, Baolin Wang wrote:

On Sat, May 9, 2020 at 7:22 AM Ming Lei [off-list ref] wrote:

quoted

Hi Sagi,

On Fri, May 08, 2020 at 03:19:45PM -0700, Sagi Grimberg wrote:

quoted

Hey Ming,

quoted

quoted

Would it make sense to elevate this flag to a request_queue flag
(QUEUE_FLAG_ALWAYS_COMMIT)?

request queue flag usually is writable, however this case just needs
one read-only flag, so I think it may be better to make it as
tagset/hctx flag.

I actually intended it to be writable.

quoted

quoted

I'm thinking of a possibility that an I/O scheduler may be used
to activate this functionality rather than having the driver set
it necessarily...

Could you explain a bit why I/O scheduler should activate this
functionality?

Sure, I've recently seen some academic work showing the benefits
of batching in tcp/ip based block drivers. The problem with the
approaches taken is that I/O scheduling is exercised deep down in the
driver, which is not the direction I'd like to go if we are want
to adopt some of the batching concepts.

I spent some (limited) time thinking about this, and it seems to
me that there is an opportunity to implement this as a dedicated
I/O scheduler, and tie it to driver specific LLD stack optimizations
(net-stack for example) relying on the commit_rq/bd->last hints.

When scanning the scheduler code, I noticed exactly the phenomenon that
this patchset is attempting to solve and Christoph referred me to it.
Now I'm thinking if we can extend this batching optimization for both
use-cases.

Got it, thanks for the sharing.

quoted

quoted

batching submission may be good for some drivers, and currently
we only do it in limited way. One reason is that there is extra
cost for full batching submission, such as this patch requires
one extra .commit_rqs() for each dispatch, and lock is often needed
in this callback.

That is not necessarily the case at all.

So far, all in-tree .commit_rqs() implementation requires lock.

quoted

quoted

IMO it can be a win for some slow driver or device, but may cause
a little performance drop for fast driver/device especially in workload
of not-batching submission.

You're mostly correct. This is exactly why an I/O scheduler may be
applicable here IMO. Mostly because I/O schedulers tend to optimize for
something specific and always present tradeoffs. Users need to
understand what they are optimizing for.

Hence I'd say this functionality can definitely be available to an I/O
scheduler should one exist.

I guess it is just that there can be multiple requests available from
scheduler queue. Actually it can be so for other non-nvme drivers in
case of none, such as SCSI.

Another way is to use one per-task list(such as plug list) to hold the
requests for dispatch, then every drivers may see real .last flag, so they
may get chance for optimizing batch queuing. I will think about the
idea further and see if it is really doable.

How about my RFC v1 patch set[1], which allows dispatching more than
one request from the scheduler to support batch requests?

[1]
https://lore.kernel.org/patchwork/patch/1210034/
https://lore.kernel.org/patchwork/patch/1210035/

Basically, my idea is to dequeue request one by one, and for each
dequeued request:

- we try to get a budget and driver tag, if both succeed, add the
request to one per-task list which can be stored in stack variable,
then continue to dequeue more request

- if either budget or driver tag can't be allocated for this request,
marks the last request in the per-task list as .last, and send the
batching requests stored in the list to LLD

- when queueing batching requests to LLD, if one request isn't queued
to driver successfully, calling .commit_rqs() like before, meantime
adding the remained requests in the per-task list back to scheduler
queue or hctx->dispatch.

One issue is that this way might degrade sequential IO performance if
the LLD just tells queue busy to blk-mq via return value of .queue_rq(),
so I guess we still may need one flag, such as BLK_MQ_F_BATCHING_SUBMISSION.


thanks, 
Ming