Install NVSHMEM on AWS EC2 instance with EFA

Accelerated Computing GPU-Accelerated Libraries
1

Hi,

I’m trying to install NVSHMEM on AWS p4d.24xlarge nodes (8xA100 per node). I notice that NVSHMEM is supported on GPUs connected via Amazon EFA, and I want to use NVSHMEM to perform inter-node communication. However, I cannot find any instructions on what optional software requirements listed here needs to be installed on my instances , and what environment variables should be set to compile NVSHMEM. Could anyone provide guidance? Thank you.

2

NVSHMEM enables support for Amazon EFA using libfabric transport. So you would need to make sure that your p4d instances have the libfabric (1.15.0.0 or later) and dependent EFA provider drivers installed. You would need to set that environment variables specified in step-7 under “building and installing NVSHMEM”: Installation — nvshmem 2.10.1 documentation