Improving CPU Performance Through Dynamic GPU Access Throttling in CPU-GPU Heterogeneous Processors

Abstract

Heterogeneous chip-multiprocessors with integrated CPU and GPU cores on the same die allow sharing of critical memory system resources among the applications executing on the twotypes of cores. In this paper, we explore memory system management driven by the quality of service (QoS) requirement of the GPU applications executing simultaneously with CPUapplications in such heterogeneous platforms. Our proposal dynamically estimates the level of QoS (e.g., frame rate in 3D scene rendering) of the GPU application. Unlike the priorproposals, our algorithm does not require any profile information and does not assume tile-based deferred rendering. If the estimated quality of service meets the minimum acceptable QoS level, our proposal employs a light-weight mechanism to dynamically adjust the GPU memory access rate so that the GPU is able to just meet the required QoS level. This frees up memory system resources which can be shifted to the co-running CPU applications. Detailed simulations done on a heterogeneous chip-multiprocessor with one GPU and four CPU cores running heterogeneous mixes of DirectX, OpenGL, and CPU applications show that our proposal improves the CPU performance by 18% on average.