Exploring CPU-GPU Coherence
University of Wisconsin – Madison, Department of Computer Sciences
University of Wisconsin – Madison, Technical report CS757, 2015
@article{subramanian2015exploring,
title={Exploring CPU-GPU Coherence},
author={Subramanian, Gokul and Thakker, Urmish and Haria, Swapnil and Shukla, Rohit and Lin, Han},
year={2015}
}
AMD, ARM and other members of the Heterogeneous Systems Architecture Foundation are focusing on integrated CPU-GPU systems with shared memory, to improve the programmability of heterogeneous systems. Such integration is also necessary to eliminate the energy and latency costs associated with conventional heterogeneous computation. This work investigates the relevance of CPU-GPU coherence for current heterogeneous workloads. We propose certain modifications to Heterogeneous Systems Coherence (HSC), one of the few feasible coherence schemes aimed at such systems, to simplify the hardware overheads and improve performance. Furthermore, we evaluate upcoming workloads like Sirius Suite to understand the application structure needed to unlock the potential of CPU-GPU coherence. Our experiments show that while integrated CPU-GPU systems with coherent shared memory improves application performance, current coherence protocols are infeasible due to the high coherence bandwidth consumed by the GPU. Our modified HSC design uses region coherence to filter GPU-side coherence requests by upto 40%.
June 8, 2015 by hgpu
