[Omp] Automated tuning of omp pragmas?

[Christian Convey]

Thanks very much Marcel.

On 5/29/07, Marcel Beemster <marcel at ace.nl> wrote:
> Christian,
>
> You wrote:
> > > > I've noticed that OpenMP offers a number of pragmas with tunable
> > > > parameters such as chunk-size.  I imagine that in a program with many
> > > > OpenMP sections, this presents a very large tuning space.
> > > >
> > > > Has anyone looked at using an automated method (i.e. a stochastic
> > > > search) to search through the tuning space to look for values that
> > > > perform well on a given target system?
>
> Can't help you directly, but I assume you refer to the
> tuning of the "schedule" clause as described in Table 2.3 on
> page 36 of the OpenMP spec.
>
> On the surface it appears to me that for each parallel
> section the schedule clause is independent (ignoring nested
> parallelism). In that case the parameters of the search
> space are independent and the search space is not that big.
> You can tick off the parallel sections one by one, keeping
> complexity low.
>
> However on an architecture where cacheing has a large impact
> on the performance you have to look at the data
> mapping between the iteration space and the individual
> processors. Preferrably you want to keep the data allocation
> similar from one parallel section to the next. This implies
> a coupling between adjacent (at run-time) parallel sections
> and then indeed the parameters are not independent any more.
>
> In my experience, it helps to have as much as possible
> information as input for an optimization algorithm. In this
> case the facts that adjacent parallel sections are coupled
> and that there is usually, though not necessarily, a fixed
> mapping from the iteration space (the i parameter) to the
> data allocation on the processors.
>
> Other than that, any optimization algorithm can be used, it
> does not have to be OpenMP specific.
>
> Best,
>         Marcel
>
> --
> Dr. Marcel Beemster, Senior Software Engineer, marcel at ace.nl,www.ace.nl
> Associated Compiler Experts bv. Amsterdam, Netherlands. +31 20 6646416.
> -----------------------------------------------------------------------
> This e-mail and any  files transmitted  with it are  confidential.  Any
> technical information contained herein is supplied as-is, and no rights
> can be  derived therefrom.  If you have received this message in error,
> please notify  the sender by reply  e-mail immediately,  and delete the
> message and all copies thereof.
>
>
>