Re: sched_rt_period_timer causing large latencies

From: Mike Galbraith
Date: Thu Apr 05 2018 - 04:11:53 EST


On Thu, 2018-04-05 at 17:44 +1000, Nicholas Piggin wrote:
>
> > My method of dealing with the throttle beast from hell for ~big box RT
> > is to stomp it flat during boot, as otherwise jitter is awful.
>
> How do you stomp it flat?

With a size 12 boot originally from SGI. Their extra hairy beasts
tended to grind themselves to fine powder.

---
Documentation/admin-guide/kernel-parameters.txt | 2 ++
kernel/sched/core.c | 11 +++++++++++
2 files changed, 13 insertions(+)

--- a/Documentation/admin-guide/kernel-parameters.txt
+++ b/Documentation/admin-guide/kernel-parameters.txt
@@ -2788,6 +2788,8 @@
noresume [SWSUSP] Disables resume and restores original swap
space.

+ nortsched [KNL] Disables realtime task bandwidth control.
+
no-scroll [VGA] Disables scrollback.
This is required for the Braillex ib80-piezo Braille
reader made by F.H. Papenmeier (Germany).
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -7091,3 +7091,14 @@ const u32 sched_prio_to_wmult[40] = {
};

#undef CREATE_TRACE_POINTS
+
+/*
+ * Bootline option to disable sched_rt_runtime.
+ */
+static int __init parse_nortsched(char *arg)
+{
+ sysctl_sched_rt_runtime = -1;
+ return 0;
+}
+early_param("nortsched", parse_nortsched);
+

> If there could be some basic detection for throttling on a per-CPU
> basis before invoking the best from hell, would that be possible to
> move this out of normal paths?

It's one beast mauling the entire box, that's the problem. I just
consider it a debugging tool, and turn it off (if it kicks in, the RT
game was already over). RT disables RT group scheduling, so I don't
have to care about that evilness.

> Alternatively I think most/all of the RT runtime is coming from
> kernel threads, could they be exempted so this doesn't hit? Although
> it would be sad to suddenly inflict these latencies on someone who
> starts up an RT program :)

Yup, very sad. It should probably become percpu.

-Mike