Re: [PATCH tip/sched/core v2] sched/rt: Simplify the IPI rt balancing logic

[Steven Rostedt]

On Fri, 5 May 2017 13:05:29 +0200
Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:

> On Thu, May 04, 2017 at 03:03:55PM -0400, Steven Rostedt wrote:
> > My test case is two fold. It basically just involves running rteval.
> > 
> > One is to run it on latest mainline to make sure it doesn't crash. The
> > other is to backport it to the latest -rt patch, and see how well it
> > helps with latency.
> > 
> > To get rteval:
> > 
> >  $ git clone git://git.kernel.org/pub/scm/linux/kernel/git/clrkwllms/rteval.git
> >  $ cd rteval
> >  $ git checkout origin/v2/master  
> 
> Blergh, that thing wants a gazillion things installed.

Blame Clark and friends ;-)

> 
> Can't I run something simple like rt-migrate-test with some arguments to
> stress the box out?

Actually what rteval does is basically 3 things. It runs cyclictest,
hackbench in a loop and a kernel build in a loop.

Note, rteval binds the the hackbench and kernel builds to nodes. That
is, if you have 4 nodes, it will run four instances of loops of both
hackbench and kernel builds in each of the nodes. This is because
hackbench and access to the filesystem across nodes with a stress test
can cause exorbitant latency due to cross node memory access on spin
locks.

I usually run cyclictest with:

 cyclictest -p80 -i250 -n -a -t -q -d 0

Although I think rteval does it slightly different. Like adding --numa
to it.

Clark, want to explain more?

-- Steve