Extending CDs to CUDA Programs

Summary

GPUs pack massive compute power within a node, which makes node-level resilience important. In this project I extended Containment Domains to CUDA programs, developing a mechanism to contain GPU errors to the GPU device itself — without requiring host-side intervention such as host-driven rollback.

Links

• Legacy project page: kyushick.github.io/project/cudacd/
• Related project: Containment Domains (CDs)