Re: [PATCH] Preemption Latency Measurement Tool

[Robert Love <rml@tech9.net>]

On Thu, 2001-09-20 at 13:37, Roger Larsson wrote:

> > Worst 20 latency times of 4261 measured in this period.
> >   usec      cause     mask   start line/file      address   end line/file
> >   4617   reacqBKL        0  1375/sched.c         c0114d94  1381/sched.c
>
> This is fantastic! It REALLY is!
> When we started with the low latency work we aimed at 10 ms.
> (in all situations, not only when running dbench... but still)

Yes it really is, especially noting that that 4.6ms lock is the
_longest_ held lock on the system.

I am seeing 90% of the reported locks under 15ms, and this means that
almost all the locks on the system are even less.

However, I am also seeing some stray 20-50ms and even 60-70ms latencies
and those bother me.  I am looking into another solution, perhaps
conditional scheduling for now.

> Lets see - no swap used? - not swapped out
> But with priority altered - it is unlikely that it would not be scheduled
> in for such a long time.
> 
> Might it be that the disk is busy to handle dbench requests. 16 threads ->
> 16 read and several (async) write requests at different disk locations in
> queue - make it 20. Seek time 10 ms => queue length 200 ms... probable???
> Do you have more than one disk? Try to run dbench on one and the player from
> the other.

Good idea.

-- 
Robert M. Love
rml at ufl.edu
rml at tech9.net