Re: [PATCH] io_uring: reduce latency by reissueing the operation

[Pavel Begunkov]

On 6/10/21 6:56 PM, Olivier Langlois wrote:
> On Thu, 2021-06-10 at 16:51 +0100, Pavel Begunkov wrote:
>> Right, but it still stalls other requests and IIRC there are people
>> not liking the syscall already taking too long. Consider
>> io_req_task_queue(), adds more overhead but will delay execution
>> to the syscall exit.
>>
>> In any case, would be great to have numbers, e.g. to see if
>> io_req_task_queue() is good enough, how often your problem
>> takes places and how much it gives us.
>>
> I will get you more more data later but I did run a fast test that
> lasted 81 seconds with a single TCP connection.
> 
> The # of times that the sqe got reissued is 57.
> 
> I'll intrumentalize a bit the code to answer the following questions:
> 
> 1. What is the ratio of reissued read sqe/total read sqe
> 2. Average exec time of __io_queue_sqe() for a read sqe when data is
> already available vs avg exec time when sqe is reissued
> 3. average exec time when the sqe is pushed to async when it could have
> been reissued.
> 
> With that info, I think that we will be in better position to evaluate
> whether or not the patch is good or not.
> 
> Can you think of other numbers that would be useful to know to evaluate
> the patch performance?

If throughput + latency (avg + several nines) are better (or any
other measurable improvement), it's a good enough argument to me,
but not sure what test case you're looking at. Single threaded?
Does it saturate your CPU?

-- 
Pavel Begunkov