Messages in this thread | | | Date | Wed, 2 Dec 2009 19:53:28 +0100 | From | Nick Piggin <> | Subject | some recent scheduler patches |
| |
Hi,
I'm just been looking through some of the recent scheduler patches while bisecting something...
83f54960c11a14942ab00b54c51e91906b9d8235: sched: for_each_domain() vs RCU I don't see what is the step that causes FAIL? sched-domains code is using synchronize_sched(), so that part should be safe.
And b8a543ea5a5896830a9969bacfd047f9d15940b2... This changelog is not correct to start with. The _idx stuff does not just shift the time that balancing decisions are made, it damps balancing choices to be more conservative if they might have been wrong over more than a single instant sample.
And secondly there is no reason give for the change. Ditto for a lot of other tuning changes really. Not that there was always exact reasons for every single one of the defaults I found, but they a) always tried to get reasonable performance with as conservative balancing as possible (ie. so 2 different tunings with no distinguishable difference then tuning that result in fewer task movements would be preferred). And b) they were relatively well tested.
Are people really not reporting enough regressions against CFS that it is time to just tweak things?
| |