On Wed, 07 Sep 2022 09:01:53 -0700 Tim Chen <tim.c.chen@xxxxxxxxxxxxxxx> wrote:
> On Thu, 2022-09-08 at 01:25 +0800, Jiebin Sun wrote:
> > The msg_bytes and msg_hdrs atomic counters are frequently
> > updated when IPC msg queue is in heavy use, causing heavy
> > cache bounce and overhead. Change them to percpu_counter
> > greatly improve the performance. Since there is one percpu
> > struct per namespace, additional memory cost is minimal.
> > Reading of the count done in msgctl call, which is infrequent.
> > So the need to sum up the counts in each CPU is infrequent.
> >
> >
> > Apply the patch and test the pts/stress-ng-1.4.0
> > -- system v message passing (160 threads).
> >
> > Score gain: 3.17x
> >
> >
> ...
> >
> > +/* large batch size could reduce the times to sum up percpu counter */
> > +#define MSG_PERCPU_COUNTER_BATCH 1024
> > +
>
> Jiebin,
>
> 1024 is a small size (1/4 page).
> The local per cpu counter could overflow to the gloabal count quickly
> if it is limited to this size, since our count tracks msg size.
>
> I'll suggest something larger, say 8*1024*1024, about
> 8MB to accommodate about 2 large page worth of data. Maybe that
> will further improve throughput on stress-ng by reducing contention
> on adding to the global count.
>
I think this concept of a percpu_counter_add() which is massively
biased to the write side and with very rare reading is a legitimate
use-case. Perhaps it should become an addition to the formal interface.
Something like
/*
* comment goes here
*/
static inline void percpu_counter_add_local(struct percpu_counter *fbc,
s64 amount)
{
percpu_counter_add_batch(fbc, amount, INT_MAX);
}
and percpu_counter_sub_local(), I guess.
The only instance I can see is
block/blk-cgroup-rwstat.h:blkg_rwstat_add() which is using INT_MAX/2
because it always uses percpu_counter_sum_positive() on the read side.
But that makes two!
