> > @@ -2378,7 +2378,9 @@ static int kswapd(void *p)
> > */
> > if (!sleeping_prematurely(pgdat, order, remaining)) {
> > trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
> > + enable_pgdat_percpu_threshold(pgdat);
> > schedule();
> > + disable_pgdat_percpu_threshold(pgdat);
>
> If we have 4096 cpus, max drift = 125x4096x4096 ~= 2GB. It is higher than zone watermark.
> Then, such sysmtem can makes memory exshost before kswap call disable_pgdat_percpu_threshold().
>
> Hmmm....
> This seems fundamental problem. current our zone watermark and per-cpu stat threshold have completely
> unbalanced definition.
>
> zone watermak: very few (few mega bytes)
> propotional sqrt(mem)
> no propotional nr-cpus
>
> per-cpu stat threshold: relatively large (desktop: few mega bytes, server ~50MB, SGI 2GB ;-)
> propotional log(mem)
> propotional log(nr-cpus)
>
> It mean, much cpus break watermark assumption.....
I've tryied to implement different patch.
The key idea is, max-drift is very small value if the system don't have
>1024CPU.
three model case:
Case1: typical desktop
CPU: 2
MEM: 2GB
max-drift = 2 x log2(2) x log2(2x1024/128) x 2 = 40 pages
Case2: relatively large server
CPU: 64
MEM: 8GBx4 (=32GB)
max-drift = 2 x log2(64) x log2(8x1024/128) x 64 = 6272 pages = 24.5MB
Case3: ultimately big server
CPU: 2048
MEM: 64GBx256 (=16TB)
max-drift = 125 x 2048 = 256000 pages = 1000MB
So, I think we can accept 20MB memory waste for good performance. but can't
accept 1000MB waste ;-)
Fortunatelly, >1000CPU machine are always used for HPC in nowadays. then,
zone_page_state_snapshot() overhead isn't so big matter on it.
So, My idea is,
Case1 and Case2: reserve max-drift pages at first.
Case3: using zone_page_state_snapshot()
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>