Messages in this thread | | | Date | Fri, 8 Jun 2012 05:13:34 -0700 | From | Anton Vorontsov <> | Subject | Re: [PATCH 2/5] vmevent: Convert from deferred timer to deferred work |
| |
On Fri, Jun 08, 2012 at 11:03:29AM +0000, leonid.moiseichuk@nokia.com wrote: > > -----Original Message----- > > From: ext Anton Vorontsov [mailto:cbouatmailru@gmail.com] > > Sent: 08 June, 2012 13:35 > ... > > > Context switches, parsing, activity in userspace even memory situation is > > not changed. > > > > Sure, there is some additional overhead. I'm just saying that it is not drastic. It > > would be like 100 sprintfs + 100 sscanfs + 2 context switches? Well, it is > > unfortunate... but come on, today's phones are running X11 and Java. :-) > > Vmstat generation is not so trivial. Meminfo has even higher overhead. I just checked generation time using idling device and open/read test: > - vmstat min 30, avg 94 max 2746 uSeconds > - meminfo min 30, average 65 max 15961 uSeconds > > In comparison /proc/version for the same conditions: min 30, average 41, max 1505 uSeconds
Hm. I would expect that avg value for meminfo will be much worse than vmstat (meminfo grabs some locks).
OK, if we consider 100ms interval, then this would be like 0.1% overhead? Not great, but still better than memcg:
http://lkml.org/lkml/2011/12/21/487
:-)
Personally? I'm all for saving these 0.1% tho, I'm all for vmevent. But, for example, it's still broken for SMP as it is costly to update vm_stat. And I see no way to fix this.
So, I guess the right approach would be to find ways to not depend on frequent vm_stat updates (and thus reads).
userland deferred timers (and infrequent reads from vmstat) + "userland vm pressure notifications" looks promising for the userland solution.
For in-kernel solution it is all the same, a deferred timer that reads vm_stat occasionally (no pressure case) + in-kernel shrinker notifications for fast reaction under pressure.
> > > In kernel space you can use sliding timer (increasing interval) + shinker. > > > > Well, w/ Minchan's idea, we can get shrinker notifications into the userland, > > so the sliding timer thing would be still possible. > > Only as a post-schrinker actions. In case of memory stressing or > close-to-stressing conditions shrinkers called very often, I saw up to > 50 times per second.
Well, yes. But in userland you would just poll/select on the shrinker notification fd, you won't get more than you can (or want to) process.
-- Anton Vorontsov Email: cbouatmailru@gmail.com
| |