> Running numa_test on my systems I find a lot of variation (as much
> as 10%) between identical runs. Are all of these results averaged
> over multiple runs? Any idea why the hackbench times vary so widely?
No, they are not averaged. I often made 2-3 runs and they usually didn't
differ significantly (within 2-3% of average user time), so I just took
one. Would be cool to have a script for averaging several results, I'll
make that if I get some time, but then we can't look at the distribution
over the nodes. And that's what I find most useful for understanding
what's going on.
> Looking at the results for my scheduler, it is real obvious what the
> algorithm is that I use for selecting a cpu for a process. This has
> not been quite as apparent on my systems, probably due to other tasks
> running in the background. The code in your scheduler that causes a
> different node to start the search for a cpu really pays off here.
Yes, but it's not perfect, as it cannot predict how long the processes
will live. Also it takes into account every running task, no matter
whether it's just a shortly running migration_thread or something else.
A running average might be better, but I have no experience with that.
Any ideas?
Regarding the hackbench times: it forks 20 tasks which fork 20 tasks
each, these are sending messages to each other. Depending on the load
balancing (which is not really predictible) it could happen that you end
up with bad distributions of senders/receivers sometimes.
Best regards,
Erich
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/