High-performance computing (HPC) has expanded beyond traditional on-premise systems, increasing the need to secure sensitive data and proprietary algorithms. Confidential Computing addresses this need by securing data in use through hardware-based Trusted Execution Environments (TEEs). Combining HPC with Confidential Computing can facilitate broader and more secure access to HPC resources, particularly in cloud environments. However, the additional confidentiality added by Confidential Computing comes at a performance cost.

This thesis extends previous research on the performance impact of AMD Secure Encrypted Virtualization-Secure Nested Paging (SEV-SNP) on HPC workloads. It provides new benchmarking results for workloads utilizing Message Passing Interface (MPI) and explores strategies to mitigate the performance overhead introduced by AMD SEV-SNP.

The results show that AMD SEV-SNP's baseline overhead can reach up to 13%. Profiling reveals decreased translation lookaside buffer (TLB) utilization, increased cache misses, and both higher frequency and longer resolve time for VMEXIT events. Optimization strategies such as the use of huge pages and paravirtualization demonstrate potential methods for reducing this overhead and improving performance.