Hi,
Here's a first few patches that provide pinned task accounting; they are boot
tested only so far.
I don't think this is enough to address Artem's regression, but its a
foundation to add some more smarts. In particular we should make it harder
still to migrate well placed tasks away due to these pinned tasks.
Rostedt, Juri, please double check what I did to your resp. set_cpus_allowed
methods.
---
include/linux/kthread.h | 1 +
include/linux/sched.h | 7 -----
kernel/kthread.c | 20 ++++++++++++---
kernel/sched/core.c | 66 ++++++++++++++++++++++++++++++++++++++++++------
kernel/sched/deadline.c | 35 ++-----------------------
kernel/sched/fair.c | 43 +++++++++++++++++++++++++------
kernel/sched/idle_task.c | 1 +
kernel/sched/rt.c | 41 +-----------------------------
kernel/sched/sched.h | 3 +++
kernel/sched/stop_task.c | 1 +
kernel/workqueue.c | 11 ++------
11 files changed, 121 insertions(+), 108 deletions(-)
On Fri, 2015-05-15 at 17:43 +0200, Peter Zijlstra wrote:
> Hi,
>
> Here's a first few patches that provide pinned task accounting; they are boot
> tested only so far.
>
> I don't think this is enough to address Artem's regression, but its a
> foundation to add some more smarts. In particular we should make it harder
> still to migrate well placed tasks away due to these pinned tasks.
Hi,
the most useful thing I can do is to test your patches, which I did, and
there seem to be no difference: average server response time is still
1.4 seconds.
Artem.