Hi Chris,

On Wed, Jan 06, 2021 at 03:43:36AM +0000, Chris Down wrote:

Feng Tang writes:

quoted

One further thought is, there are quite some "BATCH" number in
kernel for perf-cpu/global data updating, maybe we can add a
global flag 'sysctl_need_accurate_stats' for
if (sysctl_need_accurate_stats)
	batch = SMALLER_BATCH
else
	batch = BIGGER_BATCH

Moving decisions like this to the system administrator is not really a
solution to the problem -- inclusion should at least be contingent on either
having "correct-ish" stats exported to userspace. Displaying broken stats to
the user -- even with a configuration knob -- is less than ideal and is
likely to confuse and confound issues in future.

I would also like to see numbers from more real-world workloads.

Sure. Roman also mentioned this. Do you have some suggestions for the
workload or benchmarks? I don't have much knowledge on this, and have
only leveraged some of 0day's benchmarking systems.

Thanks,
Feng

MEMCG_CHARGE_BATCH is certainly fairly arbitrary as-is, but if it is going
to be changed, the reason for that change and its implications (positive and
negative) for real-world workloads must be well understood, and I'm not sure
we're there yet.