Comment 25 for bug 1833322

Revision history for this message
Doug Smythies (dsmythies) wrote : Re: Consider removing irqbalance from default install on desktop images

Hi Christian,

Thank you for your reply to my post.

> I do not know the ping pong test,

A simple token passing ring, that is useful for getting the system to utilize shallow idles states. Otherwise it can be difficult to get to such shallow states on my test system, without them being timer based.
While not relevant to this thread, the test presents a challenge for the TEO (Timer Events Orientated) idle governor, and the menu governor should perform better.

> but on iperf, I think that is in the noise
> range as far as I remember.

Agreed. I am just searching for good example type tests is all.

> If you'd just re-run that as-is what is the delta
> on your test box?

Oh. it's repeatable. I just haven't got to re-testing it yet with 24.04 as my host hardware.

> let me ask, what kind of system (cpu, size, nodes, ...)
> was that.

Intel(R) Core(TM) i5-10600K CPU @ 4.10GHz
6 cores, 2 threads per core, 12 CPUs.
HWP enabled (A.K.A. Intel Speed Shift)
intel_pstate CPU frequency driver
powersave governor
teo idle governor (but menu used below)
No throttling involved, ever.

> Your results show no change or minimal
> degradation while at the same time losing
> a bit of power. Have you also had a chance
> to try the powerthresh argument
> that Steve mentioned above?

I was just searching for a good test, and since that post I did find a really good one (not reported here). However, that was with 20.04 which has an old version of irqbalance. So I made my system dual boot adding 24.04. That same test is now not good at all for showing any differences.
And yes, also tested with the powerthresh argument.
Just for completeness, the attached graph shows processor package power and the only other graph that had some slight signal above the noise, the "idle state 1 was to deep" graph.
The test: 6 ping pong pairs, with almost no work done at each stop, 300 million loops. About 27 minutes.
Legend:
irqb-menu-disable: irqbalance disabled, menu governor.
irqb-menu-enable-1: irqbalance enabled with powerthresh=1, menu governor.
irqb-menu-enable: irqbalance enabled, menu governor.

Power: see graph, same for all.
irqbalance disabled: 5.1854 uSec/loop
irqbalance enabled powerthresh=1: 5.1966 uSec/loop
irqbalance enabled: 5.1817 uSec/loop