Does anyone have experience to share using the Intel OpenCL SDK on HPC Clusters?
Here are two motivations for my question:
1) My university department has a 192-node cluster that meets the installation requirements of the Intel OpenCL SDK. The SYSADMIN in charge of it is reasonably willing to try the installation. Is this going to be easy or difficult? Who has blazed this trail before?
2) I really want the 'big' government research machines to support OpenCL as a default configuration -- a staple capability like having a C++ and fortran compiler. So having success stories and shared experience would make an argument to do so more convincing.
I develop software for scientific computing and find the OpenCL programming model to work very well for my application, which is basically computational fluid dynamics. I want to move my code to a larger system, but have found no support through the HPC resources I am aware of. I generally work with NERSC, but NCSA and other government machines are an option.
I am aware of the NVIDIA GPU clusters. I can execute my code on those machines using OpenCL, but since my code uses OpenCL for almost all of the computation, the CPU is way underutilized. I want hybryid GPU+CPU OpenCL support and would be quite happy with OpenCL on CPU-only machines. My algorithms, afterall, perform at about the same level on modern GPU and CPU.