high performance computing on graphics processing units: hgpu.org

Trends in high performance computing are bringing increased heterogeneity among the computational resources within a single machine. The heterogeneous CPU/GPU platforms, however, exacerbate resilience problems faced by current large-scale systems. How to design efficient resilience strategies is critical for the wider adoption of heterogeneous platforms for future exascale systems. The conventional resilience strategy for GPU brings significant performance and power overhead, because they employ a one-size-fits-all approach to enforce uniform data protection. In addition, the isolation between CPU and GPU protection loses potential optimization opportunities provided by the heterogeneous CPU/GPU platforms. In this paper, we explore the viability of using an application-driven CPU/GPU cooperative method to detect faults occurred on GPU global memory. By selectively protecting application-critical data and leveraging time and space redundancy in CPU to detect faults, we bring only 2.2% performance overhead while capturing more than 90% errors that cause incorrect application results.