Re: SD_LOAD_BALANCE

2020-10-12 Thread Julia Lawall
On Mon, 12 Oct 2020, Peter Zijlstra wrote: > On Sat, Oct 10, 2020 at 06:14:23PM +0200, Julia Lawall wrote: > > Prior to v5.8 on my machine this was a rare event, because there were not > > many of these background processes. But in v5.8, the default governor for > > Intel machines without the

Re: SD_LOAD_BALANCE

2020-10-12 Thread Peter Zijlstra
On Sat, Oct 10, 2020 at 06:14:23PM +0200, Julia Lawall wrote: > Prior to v5.8 on my machine this was a rare event, because there were not > many of these background processes. But in v5.8, the default governor for > Intel machines without the HWP feature was changed from intel_pstate to > intel_cp

Re: SD_LOAD_BALANCE

2020-10-12 Thread Julia Lawall
On Mon, 12 Oct 2020, Vincent Guittot wrote: > On Mon, 12 Oct 2020 at 12:34, Julia Lawall wrote: > > > > > > Would it be useful to always check whether prev is idle, perhaps in > > > > wake_affine_idle or perhaps in select_idle_sibling? > > > > > > Yes, that would make sense to add a condition

Re: SD_LOAD_BALANCE

2020-10-12 Thread Vincent Guittot
On Mon, 12 Oct 2020 at 12:34, Julia Lawall wrote: > > > > Would it be useful to always check whether prev is idle, perhaps in > > > wake_affine_idle or perhaps in select_idle_sibling? > > > > Yes, that would make sense to add a condition in wake_affine_idle to > > return prev if this cpu is not i

Re: SD_LOAD_BALANCE

2020-10-12 Thread Julia Lawall
> > Would it be useful to always check whether prev is idle, perhaps in > > wake_affine_idle or perhaps in select_idle_sibling? > > Yes, that would make sense to add a condition in wake_affine_idle to > return prev if this cpu is not idle (or about to become idle) The case where this cpu is idle

Re: SD_LOAD_BALANCE

2020-10-12 Thread Vincent Guittot
Hi Julia, On Sat, 10 Oct 2020 at 18:14, Julia Lawall wrote: > > Hello, > > Previously, I was wondering about why starting in Linux v5.8 my unblocking > threads were moving to different sockets more frequently than in previous > releases. Now, I think that I have found the reason. > > The first i

Re: SD_LOAD_BALANCE

2020-10-10 Thread Julia Lawall
Hello, Previously, I was wondering about why starting in Linux v5.8 my unblocking threads were moving to different sockets more frequently than in previous releases. Now, I think that I have found the reason. The first issue is the change from runnable load average to load average in computing w

Re: SD_LOAD_BALANCE

2020-09-03 Thread Julia Lawall
On Thu, 3 Sep 2020, Valentin Schneider wrote: > > Hi Julia, > > On 03/09/20 15:09, Julia Lawall wrote: > > Uses of SD_LOAD_BALANCE were removed in commit e669ac8ab952 (first > > released in v5.8), with the comment: > > > > The SD_LOAD_BALANCE flag is set unconditionally for all domains in > > s

Re: SD_LOAD_BALANCE

2020-09-03 Thread Julia Lawall
On Thu, 3 Sep 2020, Valentin Schneider wrote: > > Hi Julia, > > On 03/09/20 15:09, Julia Lawall wrote: > > Uses of SD_LOAD_BALANCE were removed in commit e669ac8ab952 (first > > released in v5.8), with the comment: > > > > The SD_LOAD_BALANCE flag is set unconditionally for all domains in > > s

Re: SD_LOAD_BALANCE

2020-09-03 Thread Valentin Schneider
Hi Julia, On 03/09/20 15:09, Julia Lawall wrote: > Uses of SD_LOAD_BALANCE were removed in commit e669ac8ab952 (first > released in v5.8), with the comment: > > The SD_LOAD_BALANCE flag is set unconditionally for all domains in > sd_init(). > > I have the impression that this was not quite true.