[Beowulf] C-States and Latency

[Joe Landman]

On 02/20/2014 05:23 PM, Kevin Van Workum wrote:
> Does anyone have a real-world application where setting
> intel_idle.max_cstate=0 and processor.max_cstate=0 actually makes a
> difference wrt application or network latency?
>

Stability actually with kernels from 3.2-3.10 and certain motherboard 
and processor combos.  There are some *nasty* bugs in ASPM and other 
things, and we've observed crashes in cstate transitions in CPUs.  The 
only way we've discovered to prevent these hangs in cpu_idle and other 
related functions was specifically to disable most of the cstate 
transitions.

Some of these stability issues would take down a variety of kernel 
subsystems.  This led to very odd crashes where parts of the kernel were 
still working, but other parts we stuck in some work queue.  Eventually 
the whole thing would crash.

Later 3.10.+ kernels don't seem to have these issues (I never bisected 
to find the bad/good transition points though).



> --
> Kevin Van Workum, PhD
> Sabalcore Computing Inc.
> Run your code on 500 processors.
> Sign up for a free trial account.
> www.sabalcore.com <http://www.sabalcore.com>
> 877-492-8027 ext. 11
>
>
>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
>


-- 
Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics, Inc.
email: landman at scalableinformatics.com
web  : http://scalableinformatics.com
twtr : @scalableinfo
phone: +1 734 786 8423 x121
cell : +1 734 612 4615