Leopold Palomo Avellaneda
lepalom at wol.es
Thu Oct 23 10:17:17 EDT 2003
A Dijous 23 Octubre 2003 15:06, Ferdinand Mahr va escriure:
> Hi Leo,
> > Imaging that we have an aplication. A pararell aplication that doesn't
> use a
> > lot I/O operation, but intensive cpu, and some messages. Something like a
> > pure parallel app. We implement it using PVM or MPI ... MPI. And we make
> > test, and we have some result.
> > Now, we have our beowulf, with a linux kernel with OpenMosix with a patch
> > can migrate threads (light weith process, Mighsm,
> > or threads compiled with http://moss.csc.ncsu.edu/~mueller/pthreads/,
> > com from here:
> > benchmark.htm.
> > We have our program, and we change it that use threads for the paralel
> > behaviour and not MPI. And we run the same test. So, what will be better?
> > one have tested it?
> I haven't tested your special situation, but here are my thoughts about
> - Why changing an application that you already have? It costs you an
> unnecessary amount of time and money.
Ok, I just explaining an example. If I have to begin from 0, which approach
will be better?
> - Migshm seems to enable OpenMosix to migrate System V shared memory
> processes, not threads. But, "Threads created using the clone() system
> call can also be migrated using Migshm", that's what you want, right? I
> don't know how well that works, but it limits you to clone(), and I
> don't know if thats sufficient for reasonable thread programming. Still
> (as you mentioned before), you really can only write code that uses
> minimum I/O and interprocess/thread communication because of network
Yes, you are right. However, I hope than soon it will run pure threads. I have
heart that 2.6 have a lot of improvements in the thread part, but I'm not
> - Programs using PThreads don't run in parallel with OpenMosix/Migshm,
> they can only be migrated in whole.
Well, Pthreads can migrate with openMosix (not Linux Threads!), without the
patch. I have understood that.
> - If your MPI/PVM programs are well designed, they are usually really
> fast and can scale very well when CPU-bound.
The question that I comment is to make the programation of a parallel program
as a threads programation, and the rest is a job of the kernel in a cluster.
If this is avalaible, the management of the parallelism will be a job of the
SO, in a distributed machine.
> - Currently (Open)Mosix is better for load-balancing than HPC,
> especially in clusters with different hardware configurations. In HPC
> clusters, you usually have identical compute nodes.
> Hope that helps,
Yes, of course.
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
More information about the Beowulf