Low overhead online periodic testing for GPGPUs

GPGPUs are used to run a wide range of applications due to their high performance. As technology scales down, processing units become more susceptible to different types of faults due to radiations, manufacturing defects, wearout and aging. Some faults are detected and fixed before deployment, while others appear during infield operation. To address infield faults, continuous and periodic testing mechanisms are leveraged. In this paper, we propose a GPGPU-specific technique that takes advantage of a couple of GPGPU workloads characteristics to reduce the performance overhead of periodic testing. First, we show that many GPGPU workloads experience high probability of input similarity between threads of the same warp. Second, we show that GPGPU workloads have noticeable variation in the threads activity of their warps. Traditional periodic testing mechanisms, when applied to GPGPUs, fail to exploit these observations. Hence, we propose the half-SP deactivation technique to exploit these workload characteristics and reduce the performance overhead of the periodic testing in GPGPU platforms. The results show that the proposed technique can reduce the performance overhead of the testing from 29% to 8% with less than 1.8% area and power overheads.