On Wed, May 23, 2018 at 9:47 AM, jianchao.wang
[off-list ref] wrote:

Hi Omar

Thanks for your kindly response.

On 05/23/2018 04:02 AM, Omar Sandoval wrote:

quoted

On Tue, May 22, 2018 at 10:48:29PM +0800, Jianchao Wang wrote:

quoted

Currently, kyber is very unfriendly with merging. kyber depends
on ctx rq_list to do merging, however, most of time, it will not
leave any requests in ctx rq_list. This is because even if tokens
of one domain is used up, kyber will try to dispatch requests
from other domain and flush the rq_list there.

That's a great catch, I totally missed this.

This approach does end up duplicating a lot of code with the blk-mq core
even after Jens' change, so I'm curious if you tried other approaches.
One idea I had is to try the bio merge against the kqd->rqs lists. Since
that's per-queue, the locking overhead might be too high. Alternatively,

Yes, I used to make a patch as you say, try the bio merge against kqd->rqs directly.
The patch looks even simpler. However, because the khd->lock is needed every time
when try bio merge, there maybe high contending overhead on hkd->lock when cpu-hctx
mapping is not 1:1.

quoted

you could keep the software queues as-is but add our own version of
flush_busy_ctxs() that only removes requests of the domain that we want.
If one domain gets backed up, that might get messy with long iterations,
though.

Yes, I also considered this approach :)
But the long iterations on every ctx->rq_list looks really inefficient.

Right, this list can be quite long if dispatch token is used up.

You might try to introduce per-domain list into ctx directly, then 'none'
may benefit from this change too since bio merge should be done
on the per-domain list actually.


Thanks,
Ming Lei