Messages in this thread | | | Date | Thu, 07 Jun 2012 23:35:20 -0400 | From | KOSAKI Motohiro <> | Subject | Re: [PATCH 0/5] Some vmevent fixes... |
| |
(6/5/12 4:16 AM), leonid.moiseichuk@nokia.com wrote: >> -----Original Message----- >> From: penberg@gmail.com [mailto:penberg@gmail.com] On Behalf Of ext >> Pekka Enberg >> Sent: 05 June, 2012 11:02 >> To: Minchan Kim > ... >>> Next concern is that periodic timer of implementation. >>> I think it would add direct hook in vmscan.c rather than peeking raw >>> vmstat periodically by timer so we can control more fine-grained way >> without unnecessary overhead. >> >> If the hooks are clean and it doesn't hurt the !CONFIG_VMEVENT case, I'm >> completely OK with that. > > On the previous iteration hooking vm was pointed as very bad idea, so in my version I installed shrinker to handle cases when we have memory pressure. > Using deferred timer with adequate timeout (0.250 ms or larger) fully suitable for userspace and produce adequate overhead > -> by nature such API should not be 100% accurate, anyhow applications cannot handle situation as good as kernel can provide, 0.5MB space accuracy, 100ms is maximum user-space require for 64-1024MB devices.
I believe that's bad idea. In fact, An "adequate" timeout depend on hardware, not application performance tendency. Thus, applications can't know "adequate" value.
| |