Hello!
Tejun just this week posted patches that park the workqueue kthreads, which should reduce CPU-online latency under load. The easiest way to get these patches appears to be to pull them from Tejun's git tree:
git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq.git review-wq-hotplug
They pass some extremely mild rcutorture/hotplug testing, and more testing will follow. Please take a look at these and give them a go.
Thanx, Paul
On Fri, Jul 20, 2012 at 02:14:46PM -0700, Paul E. McKenney wrote:
Hello!
Tejun just this week posted patches that park the workqueue kthreads, which should reduce CPU-online latency under load. The easiest way to get these patches appears to be to pull them from Tejun's git tree:
git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq.git review-wq-hotplug
They pass some extremely mild rcutorture/hotplug testing, and more testing will follow. Please take a look at these and give them a go.
And they pass moderate rcutorture/hotplug testing in a number of SMP configurations. The UP tests failed due to an apparently unrelated issue that I am chasing down.
Thanx, Paul
Hi,
I have run some tests to measure the improvement of the cpu plug duration with Tejun's patches. The results are available on this link: https://wiki.linaro.org/WorkingGroups/PowerManagement/Doc/Hotplug#Workqueue
As a summary, we can see a reduction of the plug sequence compared to the kernel v3.5 as kthreads are parked instead of being destroyed
Regards, Vincent
On 21 July 2012 19:24, Paul E. McKenney paulmck@linux.vnet.ibm.com wrote:
On Fri, Jul 20, 2012 at 02:14:46PM -0700, Paul E. McKenney wrote:
Hello!
Tejun just this week posted patches that park the workqueue kthreads, which should reduce CPU-online latency under load. The easiest way to get these patches appears to be to pull them from Tejun's git tree:
git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq.git review-wq-hotplug
They pass some extremely mild rcutorture/hotplug testing, and more testing will follow. Please take a look at these and give them a go.
And they pass moderate rcutorture/hotplug testing in a number of SMP configurations. The UP tests failed due to an apparently unrelated issue that I am chasing down.
Thanx, Paul
Hello, Vincent,
Very cool, thank you for running this! So more than an order of magnitude improvement over your results from February, and significant improvement from this patch series compared to 3.5.
Not bad!
Thanx, Paul
On Thu, Jul 26, 2012 at 10:03:35AM +0200, Vincent Guittot wrote:
Hi,
I have run some tests to measure the improvement of the cpu plug duration with Tejun's patches. The results are available on this link: https://wiki.linaro.org/WorkingGroups/PowerManagement/Doc/Hotplug#Workqueue
As a summary, we can see a reduction of the plug sequence compared to the kernel v3.5 as kthreads are parked instead of being destroyed
Regards, Vincent
On 21 July 2012 19:24, Paul E. McKenney paulmck@linux.vnet.ibm.com wrote:
On Fri, Jul 20, 2012 at 02:14:46PM -0700, Paul E. McKenney wrote:
Hello!
Tejun just this week posted patches that park the workqueue kthreads, which should reduce CPU-online latency under load. The easiest way to get these patches appears to be to pull them from Tejun's git tree:
git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq.git review-wq-hotplug
They pass some extremely mild rcutorture/hotplug testing, and more testing will follow. Please take a look at these and give them a go.
And they pass moderate rcutorture/hotplug testing in a number of SMP configurations. The UP tests failed due to an apparently unrelated issue that I am chasing down.
Thanx, Paul
-
Paul E. McKenney -
Vincent Guittot