Performance and Aging Aware Resource Allocation for Concurrent GPU Applications Under Process Variation
Zois-Gerasimos Tasoulas, Iraklis Anagnostopoulos · IEEE Transactions on Nanotechnology · 2019
Platform heterogeneity prevails as a solution to the throughput and computational challenges imposed by parallel applications and technology scaling. Number-crunching or computationally intensive applications can achieve high performance when executed on a GPU. Although time multitasking is broadly available for GPUs, spatial multitaskingis not extensively explored. Randomly pairing concurrent applications and equally dividing resources among them can lead to poor performance and high divergence in the aging of the computational components. Low performance results from inefficient utilization of the available hardware, while unbalanced aging increases the probability of system failure. Additionally, process variation intensifies the effects on aging. In this paper, we present a methodology to efficiently allocate streaming multiprocessors for concurrent GPU applications. This paper considers process variation and takes actions both at the application and the kernel level. Experimental results show up to 23% and 14% higher throughput, compared to an aging aware and a performance/aging aware approach, respectively, while the relative delay deviation is reduced up to 5.9× and 4.9×, respectively.