cloud-images

Bug #1833322
Comment #11

Comment 11 for bug 1833322

Revision history for this message

Steve Langasek (vorlon) wrote on 2024-01-07: Re: Consider removing irqbalance from default install on desktop images

#11

Hi Christian,

I see a lot of strong opinions being given, but aside from the "don't use it in KVM" guidance which appears to be based on GCE's engineering expertise, very little evidence that irqbalance is actually a problem.

I think it's true that in the default config, irqbalance can interfere with putting CPUs into higher C states to conserve power. However, I don't see any indication of quantitative analysis showing the impact.

Recent versions of irqbalance have a '--powerthresh' argument that can be used to tell irqbalance to rebalance across fewer cores when CPU load is low, to allow some of the cores to be put into a sleep state and conserve power. My own initial testing on my desktop shows that this gets used for all of about 10 seconds at a time every few hours, before the load increases and irqbalance wakes the core back up...

I would want any decision to remove irqbalance from the desktop to be based on evidence, not conjecture. At a minimum, I think what I would like to see is output from powertop showing both power consumption and CPU idle stats over a reasonable amount of time (10 minutes?), on a representative client machine, for a 2x3 matrix of configurations:

- idle vs normal desktop load
- irqbalance disabled vs irqbalance enabled with defaults vs irqbalance enabled with IRQBALANCE_ARGS=--powerthresh=1

System should be rebooted between each of the irqbalance configurations, as I'm not sure what does or doesn't persist in the CPU config after irqbalance exits.

I am specifically not going to try to rebut the various webpages referenced here, beyond saying that there's an awful lot of these pages pointing to one other as authoritative sources on irqbalance without there actually being evidence to back them up (and a heaping spoonful of misinformation / outdated information along the way). So if we're going to make a change, there should be due diligence to demonstrate a benefit, it should not be based on Internet hype.

Hi Christian,

I think it's true that in the default config, irqbalance can interfere with putting CPUs into higher C states to conserve power.  However, I don't see any indication of quantitative analysis showing the impact.

Recent versions of irqbalance have a '--powerthresh' argument that can be used to tell irqbalance to rebalance across fewer cores when CPU load is low, to allow some of the cores to be put into a sleep state and conserve power.  My own initial testing on my desktop shows that this gets used for all of about 10 seconds at a time every few hours, before the load increases and irqbalance wakes the core back up...

I would want any decision to remove irqbalance from the desktop to be based on evidence, not conjecture.  At a minimum, I think what I would like to see is output from powertop showing both power consumption and CPU idle stats over a reasonable amount of time (10 minutes?), on a representative client machine, for a 2x3 matrix of configurations:

- idle vs normal desktop load
 - irqbalance disabled vs irqbalance enabled with defaults vs irqbalance enabled with IRQBALANCE_ARGS=--powerthresh=1

System should be rebooted between each of the irqbalance configurations, as I'm not sure what does or doesn't persist in the CPU config after irqbalance exits.

I am specifically not going to try to rebut the various webpages referenced here, beyond saying that there's an awful lot of these pages pointing to one other as authoritative sources on irqbalance without there actually being evidence to back them up (and a heaping spoonful of misinformation / outdated information along the way).  So if we're going to make a change, there should be due diligence to demonstrate a benefit, it should not be based on Internet hype.