Combining Cluster Sampling and ACE analysis to improve fault-injection based reliability evaluation of GPU-based systems

Computing capability demand has grown massively in recent years. Modern GPU chips are designed to deliver extreme performance for graphics and for data-parallel general purpose computing workloads (GPGPU computing) as well. Many GPGPU applications require high reliability, thus reliability evaluation has become a crucial step during their design. State-of-the-art techniques to assess the reliability of a system are fault injection and ACE analysis. The former can produce accurate results despite eternal time while the latter is very fast but it lacks accuracy of the results. In this paper we introduce a new sampling methodology based on cluster sampling that enables the exploitation of ACE analysis to accelerate the fault injection process. In our experiments we demonstrate that state-of-the-art fault injection techniques, generating random faults according to a uniform distribution, is outperformed by the proposed sampling technique, thus enabling several advantages in terms of accuracy and evaluation time. To quantify the introduced benefits we analyzed the micro-architecture reliability of an AMD Southern Islands GPU in presence of single bit upset affecting the vector register file for 6 benchmarks. One of the most important achievements is that considering all the benchmarks, on average, we are one order of magnitude faster/more accurate than uniform-sampling-based techniques in case of non exhaustive fault injection campaigns, while more than two orders of magnitude in case of exhaustive campaigns.

I. Introduction

Graphics Processing Units (GPUs) constitute an important part of the recently emerging computing continuum whose total market is more than two billion devices per year and whose application fields range from smartphones to mission-critical data center machines [1]. Technologies of this continuum have introduced benefits for several design parameters (i.e., performance and power consumption) but reliability remains a major concern [2]. Evaluating the reliability of GPU-based systems running complex applications is extremely challenging due to their hardware complexity. This requires complex and time consuming simulations. However, addressing GPUs reliability is necessary since GPUs are finding application in critical scenarios [3]. Accurate and fast techniques able to carefully trade-off between reliability analysis time and accuracy of the reported measurements are required to design complex GPU-