MKL and OpenMosix

[Tim_G_3]

Hello, we have set up an OpenMosix Cluster consisting of 6 Xeon PC's ( 2 CPUs each ). However, programs that use the MKL ( and threading ), are not being migrated to other machines but stay only on the physical machine where the job was started. I would like to now whether there is a solution to this problem, be it a different clustering technique other than OpenMosix ( which should be free, though, or at least low-cost), or a software solution - the programs we use are mostly available as source code, so re-programming or re-compilation can be taken into account. Thanks for any help. Tim Grune Dept. of Structural Chemistry University of Goettingen

[Intel_C_Intel]

Tim,

I believe your question goes beyond MKL in that it asks the question of how to get the binaries onto the nodes of the cluster. A similar question might be asked about how to get shared objects onto the nodes so they are available to programs that need them at time of execution.

I will ask one of the MKL team members who works on ScaLAPACK to respond to your question. We test and develop our ScaLAPACK code on several different clusters and do what you are seeking to do and can provide some suggestions.

Bruce

[Gregory_H_Intel]

I'm not familar with OpenMosix, but different cluster middleware packages tend to offer different methods for distributing binaries. As Bruce suggests, this is an issue beyond MKL- something any application programmer needs to be able to resolve on their cluster.

This boils down to a cluster middleware issue.

[Gregory_H_Intel]

I'm not familar with OpenMosix, but different cluster middleware packages tend to offer different methods for distributing binaries. As Bruce suggests, this is an issue beyond MKL- something any application programmer needs to be able to resolve on their cluster.

This boils down to a cluster middleware issue.

With

[Gregory_H_Intel]

I'm not familar with OpenMosix, but different cluster middleware packages tend to offer different methods for distributing binaries. As Bruce suggests, this is an issue beyond MKL- something any application programmer needs to be able to resolve on their cluster.

This boils down to a cluster middleware issue.

With clusters built with OSCAR, for instance, OSCAR automatically mounts the /home directory so that anything in /home is available to any node. Then MPICH commands like mpirun automatically have all the data they need. I'm told NPACI/Rocks does roughly the same thing through export/install.

Again, I don't know what mechanisms OpenMosix uses to do these kinds of things, but it must have some mechanism for launching and migrating MPI jobs, or there must be some additional software you can install that simplify this process.

- Greg