Study on Heterogeneous Queuing

Abstract

As CPU processing speed has slowed down year-on-year, heterogeneous “CPU-GPU” architectures combining multi-core CPU and GPU accelerators have become increasingly attractive. Under this backdrop, the Heterogeneous System Architecture (HSA) standard was released in 2012. New Accelerated Processing Unit (APU) architectures – AMD Kaveri and Carrizo – were released in 2014 and 2015 respectively, and are compliant with HSA. These architectures incorporate two technologies central to HSA, hUMA (heterogeneous Unified Memory Access) and hQ (heterogeneous Queuing). This paper summarizes the detailed processes of hQ by analyzing the AMDKFD kernel source code. Furthermore, this paper also presents hQ performance indexes obtained by running matrix-vector multiplications on Kaveri and Carrizo experiment platforms. The experimental results show that hQ can prevent the system from falling into kernel mode as much as possible without additional overhead. We find that compared with Kaveri, the Carrizo architecture provides better HSA performance.