Mark Hull-Richter wrote:
>
> A while back I sent out an email requesting good benchmarks for SMP
> Alpha machines. The only seemingly viable response I received was one
> recommending the SPLASH benchmark, with a request to post my results. I
> obtained and attempted to use the SPLASH2 set of programs which are
> available from www-flash.stanford.edu.
>
> Here is what I found:
>
> 1) The benchmark programs do not compile as downloaded. There is some
> problem with missing identifiers and external functions. This was done
> with the null macros supplied with the benchmark suite on a UP Alpha
> (our development platform). The null macros are supposed to nullify the
> SMP nature of the tests and allow them to be compiled for a UP test.
> This failed.
>
> 2) The only available references in the supplied material are obsolete
> email addresses for people/groups that I would have to guess are no
> longer available at Stanford.
>
> 3) The only available sets of macros that are "done" are for three
> machines, none of which matches Linux or Alpha. The closest set, that
> for SGI's IRIX, include a precompiled library with no apparent source
> and no references on how to construct an equivalent for other machines.
>
> I suppose that I could spend a lot of time and effort tracking these
> specific problems down and resolving them, and I have no doubt that I
> could. However, my charter at the moment does not allow for this kind
> of time and effort, and is unlikely to for the foreseeable future.
>
> Regards.
Hello
I think I can help you, at least about the SPLASH2 codes
I am successfully using the SPLASH2 on multiple cpu with
posix threads (tested under linux, Irix 6.4 and solaris 2.7
but should work as well with digital unix).
I'm talking about the SPLASH2 here, I've also used the SPLASH 1
in the past and they requiered a little more work to run. Most
of the SPLASH1 are included in the SPLASH2.
To do that I implemented the set of m4 macros PARMACS for the
posix threads. After doing that, it works well. I even used
this macro set for my own parallel application
(PMesa: http://www.lri.fr/~alex/PMesa )
(I've also implemented the macros for macos and windows nt).
If you are interested, I can make this available.
I'm not sure the SPLASH are really the good benchmark for the
linux scheduler because they are cpu-intensive and don't call
the kernel (to synchronise the threads) very often. For instance,
they run as fast with 2.0 and 2.2. Only if you don't take care
of keeping a thread always on the same CPU, then you misuse the
private L2 cache and the performance drops (15% loss)
But they are often used for research on (real or simulated)
parallel machines with a big number of cpus.
Alexis
--
*******************************************************
Alexis VARTANIAN - tel: (33) 1 69 15 64 07
Equipe Architectures Paralleles, LRI
Batiment 490, Universite Paris-Sud, 91405 Orsay, France
*******************************************************
-
Linux SMP list: FIRST see FAQ at http://www.irisa.fr/prive/mentre/smp-faq/
To Unsubscribe: send "unsubscribe linux-smp" to [EMAIL PROTECTED]
