Method for Using the MVAPICH2 Performance Optimization Option

You can use the environment variables within the job script to set the inter-process communication algorithm when using MVAPICH2 as an MPI library. The user can optimize the execution performance by setting the environment variables for the collective communication, which is often used in the code. The main keywords are as follows.

The MV2_INTRA environment variable is used for the intra-node MPI process communication, and the MV2_INTER environment variable is used for the inter-node MPI process communication. For MPI_Alltoall, the same algorithm is applied to both the INTRA and INTER cases. The same can be applied to routines for transmitting data of different sizes (example: MPI_Alltoallv) or non-blocking-based routines (example: MPI_Ibcast), and the setting method is shown below.

• Main environment variables for the MVAPICH2 library collective communication

MPI communication	Environment variable	Setting range
MPI_Gather	MV2_INTRA_GATHER_TUNING MV2_INTER_GATHER_TUNING	0 ~ 3
MPI_Bcast	MV2_INTRA_BCAST_TUNING MV2_INTER_BCAST_TUNING	0 ~ 10
MPI_Scatter	MV2_INTRA_SCATTER_TUNING MV2_INTER_SCATTER_TUNING	1 ~ 5
MPI_Allreduce	MV2_INTRA_ALLREDUCE_TUNING MV2_INTER_ALLREDUCE_TUNING	1 ~ 6
MPI_Reduce	MV2_INTRA_REDUCE_TUNING MV2_INTER_REDUCE_TUNING	1 ~ 6
MPI_Allgather	MV2_INTRA_ALLGATHER_TUNING MV2_INTER_ALLGATHER_TUNING	1 ~ 11
MPI_Alltoall	MV2_ALLTOALL_TUNING	0 ~ 4

※ Usage example (when creating a bash script): export MV2_INTER_ALLREDUCE_TUNING=2

※ Refer to http://mvapich.cse.ohio-state.edu/userguide/ for a detailed description of each environment variable.

2022년 9월 15일에 마지막으로 업데이트되었습니다.