Effective Utilization of CUDA Hyper-Q for Improved Power and Performance Efficiency
Authors:
- Ryan S. LuleyRLRyan S. Luley
- Qinru QiuQQQinru Qiu
- 1 May 2016
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1160-1169
Abstract
High utilization of hardware resources is the key for designing performance and power optimized GPU applications. The efficiency of applications and kernels, which do not fully utilize the GPU resources, can be improved through concurrent execution with independent kernels and/or applications. Hyper-Q enables multiple CPU threads or processes to launch work on a single GPU simultaneously for increased GPU utilization. However, without careful design, false serialization may occur due to the contention for shared hardware resources such as direct memory access (DMA) engines. In this paper, we reveal the impact of such contention on performance and assess a method for overcoming the limitation with minimal algorithmic overhead. We demonstrate a method to achieve up to 31.8% improvement in performance and 10.4%reduction in energy on average for a finite set of application tasks when maximizing GPU execution concurrency.Keywords
This publication has 9 references indexed in Scilit:
- Efficient GPU Spatial-Temporal MultitaskingIEEE Transactions on Parallel and Distributed Systems, 2015
- Symbiotic scheduling of concurrent GPU kernels for performance and energy optimizations2014
- Exploring Graphics Processing Unit (GPU) Resource Sharing Efficiency for High Performance ComputingComputers, 2013
- Improving GPGPU concurrency with elastic kernels2013
- On Improving the Performance of Multi-threaded CUDA Applications with Concurrent Kernel Execution by Kernel Reordering2012
- The case for GPGPU spatial multitasking2012
- A characterization of the Rodinia benchmark suite with comparison to contemporary CMP workloads2010
- An integrated GPU power and performance model2010
- Rodinia: A benchmark suite for heterogeneous computing2009