Device specialization in heterogeneous multi-GPU environments

Abstract

In the last few years there have been many activities towards coupling CPUs and GPUs in order to get the most from CPU-GPU heterogeneous systems. One of the main problems that prevent these systems to be exploited in a device-aware manner is the CPU-GPU communication bottleneck, which often doesn't allow to produce code more efficient than the GPU-only and the CPU-only counterparts. As a consequence, most of the heterogeneous scheduling systems treat CPUs and GPUs as homogeneous nodes, electing map-like data partitioning to employ both these processing resources.  We propose to study how the radical change in the connection between GPU, CPU and memory characterizing the APUs (Accelerated Processing Units) affect the architecture of a compiler and if it is possible to use all these computing resources in a device-aware manner. We investigate on a methodology to analyze the devices that populate heterogeneous multi-GPU systems and to classify general purpose algorithms in order to perform near-optimal control flow and data partitioning.