Re: [PATCH] schedstat-2.6.0-test5-A1 measuring process scheduling latency

[Davide Libenzi <davidel@xmailserver.org>]

On Thu, 11 Sep 2003, Rick Lindsley wrote:

> All the tweaking of the scheduler for interactive processes is still
> rather reliant on "when I run X, Y, and Z, and wiggle the mouse, it feels
> slower."  At best, we might get a more concrete "the music skips every
> now and then."  It's very hard to develop changes for an environment I
> don't have in front of me and can't ever seem to precisely duplicate.
>
> It's my contention that what we're seeing is scheduler latency: processes
> want to run but end up waiting for some period of time.  There's currently
> no way to capture that short of the descriptions above.  This patch adds
> counters at various places to measure, among other things, the scheduler
> latency.  That is, we can now measure the time processes spend waiting
> to run after being made runnable.  We can also see on average how long
> a process runs before giving up (or being kicked off) the cpu.

When I was working on it, I wrote a patch that at every context switch
used to register cycle counter timestamp, pid and other things (this per
CPU), with the number of samples configurable. This helped me a lot not
only to understand latencies, but also process migrations (usefull for
balancing thingies). There was a /dev/something where you could collect
samples and analyze them. It'd nice to have something like that for 2.6.



- Davide