Re: SCHED_ULE should not be the default

On Monday, December 12, 2011 12:06:04 pm Steve Kargl wrote:
> On Mon, Dec 12, 2011 at 04:18:35PM +0000, Bruce Cran wrote:
> > On 12/12/2011 15:51, Steve Kargl wrote:
> > >This comes up every 9 months or so, and must be approaching FAQ 
> > >status. In a HPC environment, I recommend 4BSD. Depending on the 
> > >workload, ULE can cause a severe increase in turn around time when 
> > >doing already long computations. If you have an MPI application, 
> > >simply launching greater than ncpu+1 jobs can show the problem. PS: 
> > >search the list archives for "kargl and ULE". 
> > 
> > This isn't something that can be fixed by tuning ULE? For example for 
> > desktop applications kern.sched.preempt_thresh should be set to 224 from 
> > its default. I'm wondering if the installer should ask people what the 
> > typical use will be, and tune the scheduler appropriately.
> > 
> 
> Tuning kern.sched.preempt_thresh did not seem to help for
> my workload.  My code is a classic master-slave OpenMPI
> application where the master runs on one node and all
> cpu-bound slaves are sent to a second node.  If I send
> send ncpu+1 jobs to the 2nd node with ncpu's, then 
> ncpu-1 jobs are assigned to the 1st ncpu-1 cpus.  The
> last two jobs are assigned to the ncpu'th cpu, and 
> these ping-pong on the this cpu.  AFAICT, it is a cpu
> affinity issue, where ULE is trying to keep each job
> associated with its initially assigned cpu.
> 
> While one might suggest that starting ncpu+1 jobs
> is not prudent, my example is just that.  It is an
> example showing that ULE has performance issues. 
> So, I now can start only ncpu jobs on each node
> in the cluster and send emails to all other users
> to not use those node, or use 4BSD and not worry
> about loading issues.

This is a case where 4BSD's naive algorithm will spread out the load more
evenly because all the threads are on a single, shared queue and each CPU
just grabs the head of the queue when it finishes a timeslice.  ULE always
assigns threads to a single CPU (even if they aren't pinned to a single
CPU using cpuset, etc.) and then tries to balance the load across cores
later, but I believe in this case it's rebalancer won't have anything to
really do as no matter what it does with the N+1 job it's going to be
sharing a CPU with another job.

-- 
John Baldwin