lkml.org 
[lkml]   [2019]   [Jun]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v4 0/5] x86/umwait: Enable user wait instructions
On Tue, Jun 11, 2019 at 10:37:34AM -0700, Fenghua Yu wrote:
> On Tue, Jun 11, 2019 at 11:01:45AM +0200, Peter Zijlstra wrote:
> > On Fri, Jun 07, 2019 at 03:00:32PM -0700, Fenghua Yu wrote:
> > > Today, if an application needs to wait for a very short duration
> > > they have to have spinloops. Spinloops consume more power and continue
> > > to use execution resources that could hurt its thread siblings in a core
> > > with hyperthreads. New instructions umonitor, umwait and tpause allow
> > > a low power alternative waiting at the same time could improve the HT
> > > sibling perform while giving it any power headroom. These instructions
> > > can be used in both user space and kernel space.
> > >
> > > A new MSR IA32_UMWAIT_CONTROL allows kernel to set a time limit in
> > > TSC-quanta that prevents user applications from waiting for a long time.
> > > This allows applications to yield the CPU and the user application
> > > should consider using other alternatives to wait.
> >
> > I'm confused on the purpose of this control; what do we win by limiting
> > this time?
>
> In previous patches, there is no time limit (max time is 0 which means no
> time limit).
>
> Andy Lutomirski proposed to set the time limit:
>
> https://lkml.org/lkml/2019/2/26/735
>
> "So I propose setting the timeout to either 100 microseconds or 100k
> "cycles" by default. In the event someone determines that they save
> materially more power or gets materially better performance with a
> longer timeout, we can revisit the value."
>
> Does it make sense?

You quoted exactly the wrong part of that message; Andy's concern was
with NOHZ_FULL. And I think we should preserve that concern in both the
code and Changelog introducing this limit.

\
 
 \ /
  Last update: 2019-06-17 16:19    [W:0.053 / U:6.100 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site