What's the right way to measure latency?

What's the right way to measure latency?

[Chris Worley]

I'm trying to measure latency.  Is this the right incantation for "8KB
block write latency"?

fio --name=test --numjobs=1 --bs=8k --rw=randwrite --sync=0 --direct=1
--iodepth=1 --ioengine=libaio --loops=1 --size=16091503001
--filename=/dev/sdb

... I figure I can only measure latency w/ just one thread and no
outstanding I/O's.

The results show two latency numbers: clat and slat... what are these?

Thanks,

Chris

Re: What's the right way to measure latency?

[Joshua Aune]

On Apr 26, 2009, at 7:44 AM, Chris Worley wrote:

> The results show two latency numbers: clat and slat... what are these?

slat is not there for the sync engines but exists for libaio, submit  
latency.  clat sounds like completion latency.

Josh

Re: What's the right way to measure latency?

[Jens Axboe]

On Sun, Apr 26 2009, Joshua Aune wrote:
>
> On Apr 26, 2009, at 7:44 AM, Chris Worley wrote:
>
>> The results show two latency numbers: clat and slat... what are these?
>
> slat is not there for the sync engines but exists for libaio, submit  
> latency.  clat sounds like completion latency.

It's in the HOWTO :-)

Basically slat is the time it too to enter the kernel and submit the IO
(submission latency) and clat is the time it took from submission to the
IO completed (completion latency). With the sync engines you cannot
separate the two, since when the submission syscall completes the IO has
also completed. Hence they have only clat numbers, where the async
engine has the ability to measure both.

-- 
Jens Axboe