Re: [PATCH 0/10] psi: pressure stall information for CPU, memory, and IO v2

[Singh, Balbir]

On 7/25/18 1:15 AM, Johannes Weiner wrote:
> Hi Balbir,
> 
> On Tue, Jul 24, 2018 at 07:14:02AM +1000, Balbir Singh wrote:
>> Does the mechanism scale? I am a little concerned about how frequently
>> this infrastructure is monitored/read/acted upon.
> 
> I expect most users to poll in the frequency ballpark of the running
> averages (10s, 1m, 5m). Our OOMD defaults to 5s polling of the 10s
> average; we collect the 1m average once per minute from our machines
> and cgroups to log the system/workload health trends in our fleet.
> 
> Suren has been experimenting with adaptive polling down to the
> millisecond range on Android.
> 

I think this is a bad way of doing things, polling only adds to overheads, there needs to be an event driven mechanism and the selection of the events need to happen in user space.

>> Why aren't existing mechanisms sufficient
> 
> Our existing stuff gives a lot of indication when something *may* be
> an issue, like the rate of page reclaim, the number of refaults, the
> average number of active processes, one task waiting on a resource.
> 
> But the real difference between an issue and a non-issue is how much
> it affects your overall goal of making forward progress or reacting to
> a request in time. And that's the only thing users really care
> about. It doesn't matter whether my system is doing 2314 or 6723 page
> refaults per minute, or scanned 8495 pages recently. I need to know
> whether I'm losing 1% or 20% of my time on overcommitted memory.
> 
> Delayacct is time-based, so it's a step in the right direction, but it
> doesn't aggregate tasks and CPUs into compound productivity states to
> tell you if only parts of your workload are seeing delays (which is
> often tolerable for the purpose of ensuring maximum HW utilization) or
> your system overall is not making forward progress. That aggregation
> isn't something you can do in userspace with polled delayacct data.

By aggregation you mean cgroup aggregation?

> 
>> -- why is the avg delay calculation in the kernel?
> 
> For one, as per above, most users will probably be using the standard
> averaging windows, and we already have this highly optimizd
> infrastructure from the load average. I don't see why we shouldn't use
> that instead of exporting an obscure number that requires most users
> to have an additional library or copy-paste the loadavg code.
> 
> I also mentioned the OOM killer as a likely in-kernel user of the
> pressure percentages to protect from memory livelocks out of the box,
> in which case we have to do this calculation in the kernel anyway.
> 
>> There is no talk about the overhead this introduces in general, may be
>> the details are in the patches. I'll read through them
> 
> I sent an email on benchmarks and overhead in one of the subthreads, I
> will include that information in the cover letter in v3.
> 
> https://lore.kernel.org/lkml/20180718215644.GB2838@xxxxxxxxxxx/

Thanks, I'll take a look

Balbir Singh.