Evaluating MPI Performance on SGX and Gramine
Kota Shimojima, Shinobu Miwa, Hayato Yamaki, Hiroki Honda · 2024
In this paper, we provide the first analysis of Gramine+SGX for MPI performance. For this, we constructed a small PC cluster, each node having one Xeon processor that supports SGX. We executed 16 MPI programs selected from the NAS Parallel Benchmarks and Intel MPI Benchmarks on this PC cluster, and found that the overhead of Gramine+SGX on network performance was small in many cases (i.e., within 30% for 7 of the 8 NPB programs executed on 4 nodes). This paper is an important step towards confidential computing in the field of HPC.