compiler and hardware support status for CAF running on MIC in native mode?

[Izaak_Beekman]

Is my understanding correct, that the current Beta and the upcoming 15.x release of ifort will support Coarrays on MIC (xeon phi)?

Will the compiler and hardware be able to generate code that can run on the Xeon Phi in “native” mode, rather than “offload” or “symmetric” modes? By native mode, I mean that it will be able to execute on machines like the DOE’s upcoming Cori where each instance of the SPMD code executes entirely on the xeon phi, with multiple instances (if we were talking about MPI rather than CAF, instance = MPI rank) per Xeon Phi.

TIA

[TimP]

The announcement of Cori said that it would incorporate the future KNL version of Intel(r) Xeon Phi(tm) (to support independent Phi nodes as opposed to coprocessor) so that seems too far in the future to answer all those questions.  The 15.0 beta compiler coarrays are documented as not working with the new Intel MPI 5.0 (previous MPI required on host) so it doesn't appear that such features can be expected in the imminent 15.0 compiler.  15.0 beta ifort (and recent gfortran) had option to generate AVX512 code presumably for KNL but at present that could run only on the SDE (software simulator).  I suppose Intel people would have to read the regulations carefully to see how much they can say, if anything, about what is planned for KNL.

[Steven_L_Intel1]

Version 14 already supports coarrays on MIC, including native and heterogeneous models (not offload - some images run on the Xeon Phi and some on the Xeon host.) This isn't new in version 15.

[Izaak_Beekman]

Steve Lionel (Intel) wrote:

Version 14 already supports coarrays on MIC, including native and heterogeneous models (not offload - some images run on the Xeon Phi and some on the Xeon host.) This isn't new in version 15.

Steve, thanks for the prompt reply. The heterogeneous model is what I was referring to as symmetric. I’m please to hear that there is already extant support for coarrays on Xeon Phi using the native execution model.

Can you speak to Tim’s points about system requirements to get started doing this, or point me to some good documentation? Tim mentions that it doesn’t work with the newest Intel MPI 5 implementation. Will it work with older Intel MPI?

Tim, I am aware that Cori will have independent KNL Phi nodes, which is why I am interested in native execution. Obviously the system has not been deployed yet, so I am not asking about specifics related to KNL/independent Phi nodes/Cori, but rather, if it is possible to start getting ready for Cori—and similar—using existing systems (like Babbage). To do so, one will have to use the native execution model, since, on Cori, the node will consist of only the Xeon Phi and there is nowhere else execute instances of the SPMD code. Hopefully, once such systems are deployed, Intel and/or Cray,etc. Fortran compilers will fully support coarrays on such architectures.

[Steven_L_Intel1]

Tim, I am puzzled by your response. In fact, Intel MPI 5 is the only version we support for x64 in the 15.0 compiler. In 14.0, we support Intel MPI 4.1.3. Instructions for using coarrays with MIC are in the compiler release notes.

Coincidentally, we discovered only earlier this week that the MPI redistributable files for MIC are missing from the compiler install. We will fix that as soon as we can. You can get those files from the Intel MPI redistributables package, however.

[Izaak_Beekman]

Great news, thanks for getting back to me Steve, I really appreciate it.