6.3 Module and storage environment on Taito-GPU
Taito-GPU shares the same module and storage environment as the rest of the Taito cluster system. The actual module tree is, however, optimized for GPU usage and thus different. Regarding file system directories, see more from Chapter 1.5 of Taito User Guide.
Please pay attention to your module environment. Due to dynamic linking nature of application program it is recommended that you always have exactly the same environment when running your application as when you used to compile & link it.
If you are using pure CUDA C++ with your C/C++-application, the following module setup is usually sufficient:
module purge module load gcc cuda
If you opt for directive based OpenACC approach, then the following setup usually does the job – activates PGI-compiler to interpret OpenACC directives:
module purge module load pgi cuda
In case of multi-GPU MPI application it is advisable to enable also CUDA-aware message passing. This allows the MPI implementation to directly process device pointers with no need by the user to transfer them from device to host before the transfer.
In case of OpenMPI, load:
module purge module load gcc cuda openmpi
In case of MVAPICH2, load:
module purge module load gcc cuda mvapich2
|Previous chapter||One level up||Next chapter|