lkml.org 
[lkml]   [2018]   [Nov]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [RFC PATCH v1 2/2] proc: add /proc/<pid>/thread_state
From
Date
On 2018/11/8 18:17, Peter Zijlstra wrote:
> On Thu, Nov 08, 2018 at 07:32:46AM +0100, Ingo Molnar wrote:
>>
>> * Aubrey Li <aubrey.li@intel.com> wrote:
>>
>>> Expose the per-task cpu specific thread state value, it's helpful
>>> for userland to classify and schedule the tasks by different policies
>>
>> That's pretty vague - what exactly would use this information? I'm sure
>> you have a usecase in mind - could you please describe it?
>
> Yeah, "thread_state" is a pretty terrible name for this.

task_struct has a CPU specific element "thread", I quote it to here to
create a generic interface to expose CPU specific state of the task.
Like /proc/<pid>/stat, I plan to use this "thread_state" to host any CPU
specific state, including AVX state(now only), and any other states(may come
soon). So this interface may be extended in future like the following:

#cat /proc/<pid>/thread_state
1 0 0


>
The use-case is
> detectoring which tasks use AVX3 such that a userspace component (think
> job scheduler) can cluster them together.>
> The 'problem' is that running AVX2+ code drops the max clock, because
> you light up the massive wide (and thus large area) paths.

Thanks to explain this.

>
> So maybe something like "simd_active" is a better name, dunno.

As above, maybe the name can be hidden by thread_state?

>
> Or maybe "simd_class" and we can write out 0,1,2,3 depending on the AVX
> class being used, dunno. It might make sense to look at what other arch
> SIMD stuff looks like to form this interface.
>
A task may use 1,2,3 simultaneously, as a scheduler hint, it's just cluster
or not, so 0/1 may be good enough.

Thanks,
-Aubrey

\
 
 \ /
  Last update: 2018-11-08 15:03    [W:0.074 / U:5.876 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site