On 5/16/16 5:58 PM, Paul E. McKenney wrote:
On Mon, May 16, 2016 at 12:49:41PM -0700, Santosh Shilimkar wrote:
On 5/16/2016 10:34 AM, Paul E. McKenney wrote:
On Mon, May 16, 2016 at 09:33:57AM -0700, Santosh Shilimkar wrote:
[...]
Are you running CONFIG_NO_HZ_FULL=y? If so, the problem might be that
you need more housekeeping CPUs than you currently have configured.
Yes, CONFIG_NO_HZ_FULL=y. Do you mean "CONFIG_NO_HZ_FULL_ALL=y" for
book keeping. Seems like without that clock-event code will just use
CPU0 for things like broadcasting which might become bottleneck.
This could explain connect the hrtimer_interrupt() path getting slowed
down because of book keeping bottleneck.
$cat .config | grep NO_HZ
CONFIG_NO_HZ_COMMON=y
# CONFIG_NO_HZ_IDLE is not set
CONFIG_NO_HZ_FULL=y
# CONFIG_NO_HZ_FULL_ALL is not set
# CONFIG_NO_HZ_FULL_SYSIDLE is not set
CONFIG_NO_HZ=y
# CONFIG_RCU_FAST_NO_HZ is not set
Yes, CONFIG_NO_HZ_FULL_ALL=y would give you only one CPU for all
housekeeping tasks, including the RCU grace-period kthreads. So you are
booting without any nohz_full boot parameter? You can end up with the
same problem with CONFIG_NO_HZ_FULL=y and the nohz_full boot parameter
that you can with CONFIG_NO_HZ_FULL_ALL=y.
I see. Yes, the systems are booting without nohz_full boot parameter.
Will try to add more CPUs to it & update the thread
after the verification since it takes time to reproduce the issue.
Thanks for discussion so far Paul. Its very insightful for me.
Regards,
Santosh
