Heterogeneous cluster support

[lsr]

Can both Windows and Linux versions of Intel MPI work together?

Also, is there some substitute for the lack of support for MPI-2 process spawning? We produce simulation libs which clients (including GUI-enabled) link against and we need to implement portable distributed processing, so launching via mpiexec isn't a viable option - we need the (MPI-enabled) client app to be able to launch worker MPI processes and communicate with/manage those.

Thanks.

[TimP]

Quoting - lsr

Can both Windows and Linux versions of Intel MPI work together?

I don't know what this means. The corresponding Windows and linux versions should have the same level of support for MPI, and work the same at the MPI call level. They use different underlying OS functions.
I don't see any likelihood of MPI jobs on Windows and linux communicating while running, but it might be interesting, if you would explain.

You could submit feature requests about additional MPI-2 functionality on your premier.intel.com account.

[lsr]

Quoting - tim18

I don't know what this means. The corresponding Windows and linux versions should have the same level of support for MPI, and work the same at the MPI call level. They use different underlying OS functions.
I don't see any likelihood of MPI jobs on Windows and linux communicating while running, but it might be interesting, if you would explain.

You could submit feature requests about additional MPI-2 functionality on your premier.intel.com account.

I mean a cluster of Windows and Linux machines running a single MPI job consisting of processes linked against the corresponding, platform-dependant version of Intel MPI library. Typically, I'll have a master/controller, GUI-enabled app running on Windows workstation and, say, a dozen of Linux boxes for number crunching.

Thanks for your suggestion for requesting process spawning, but I'm not Intel customer, at least not yet. I'm evaluating Intel MPI and if they really don't have some form of process management, that's a show stopper for me. I don't think I want to do IPC and I'll probably resort to using MPICH2 instead.

[TimP]

I suppose, if you're trying to implement a Windows interface to torque running on linux (just one possible guess as to what you mean), it won't matter which MPI you use. Microsoft has a near monopoly on Windows cluster job management, so they would have a big say on the required budget and capabilities.

[Dmitry_K_Intel2]

According to theIntel MPI Release notes: "theIntel MPI Library does not support heterogeneous clusters of mixed architectures and/or operating environments." If you need this feature you can submit a request on https://premier.intel.com/

[cfrieler]

Quoting - lsr

Can both Windows and Linux versions of Intel MPI work together?

Also, is there some substitute for the lack of support for MPI-2 process spawning? We produce simulation libs which clients (including GUI-enabled) link against and we need to implement portable distributed processing, so launching via mpiexec isn't a viable option - we need the (MPI-enabled) client app to be able to launch worker MPI processes and communicate with/manage those.

Thanks.

My recollection from (several years ago) was that this was possible with MPICH to some extent. We were able to run a single large job across a heterogeneous collection of systems, but it required a seperate compile for each platform and a configuration file that specified the correct binary for each node. We succeeded in getting demonstration codes and a couple simple problems distributed across the whole collection. However, we were really trying to leverage our Unix platforms and the data file incompatibilities killed the project.