RE: REGRESSION BISECTED: mq-deadline: covert to use request_queue->async_depth

[Doug Smythies]

Hi Kuai,

Thank you for your replies, and the time you took to write them.

On 2026.04.10 09:21 Yu Kuai wrote:

> Hi,
> 
> 在 2026/4/10 13:18, Doug Smythies 写道:
>>> Do you ever use the async_depth for mq-deadline?
>> No.
>
> Then the only difference is the tag wake batch is changed from 1 to 8. However,
> batch wake up is the base design of sbitmap. And I do admit tail latency is
> bad under high concurrency is probably due to wake up is unfair, this is a
> know sbitmap design flaw and not fixed yet.

Good to know.

> You could try change wake up batch back to 1. However, I don't think this is a
> good solution. And I'll suggest you to try increase nr_requests to see if tail
> Latency gets better.

I just observed what I thought was a pretty significant regression, and thought
I should bring it the attention of the stake holders.

By decreasing the maximum concurrency of my workflow from 1000 to 64 the regression
decreased to noise levels, +/- 1.5%.

Thanks,

Doug Smythies