Co-Scheduling on Fused CPU-GPU Architectures With Shared Last Level Caches
- 18 July 2018
- journal article
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 37 (11) , 2337-2347
Abstract
Fused CPU-GPU architectures integrate a CPU and general-purpose GPU on a single die. Recent fused architectures even share the last level cache (LLC) between CPU and GPU. This enables hardware-supported byte-level coherency. Thus, CPU and GPU can execute computational kernels collaboratively, but novel methods to co-schedule work are required. This paper contributes three dynamic co-scheduling methods. Two of our methods implement workers that autonomously acquire work from a common set of independent work items (similar to bag-of-tasks scheduling). The third method, host-side profiling , uses a fraction of the total work of a kernel to determine a ratio of how to distribute work to CPU and GPU based on profiling. The resulting ratio is used for the following executions of the same kernel. Our methods are realized using OpenCL 2.0, which introduces fine-grained shared virtual memory (SVM) to allocate coherent memory between CPU and GPU. We port the Rodinia Benchmark Suite, a standard suite for heterogeneous computing, to fine-grained SVM and fused CPU-GPU architectures ( Rodinia-SVM ). We evaluate the overhead of fine-grained SVM and analyze the suitability of OpenCL 2.0’s new features for co-scheduling. Our host-side profiling method performs competitively to the optimal choice of executing kernels either on CPU or GPU (hypothetical xor-Oracle). On average, it achieves 97% of xor-Oracle’s performance and a $1.43\times$ speedup over using the GPU alone (standard in Rodinia). We show, however, that in most cases it is not beneficial to split the work of a kernel between CPU and GPU compared to exclusively running it on the most suitable single compute device. For a fixed amount of work per device, cache-related stalls can increase by up to $1.75\times$ when both devices are used in parallel instead of exclusively while cache misses remain the same. Thus, not the cost of cache conflicts, but inefficient cache coherence is a major performance bottleneck for current fused CPU-GPU Intel architectures with shared LLC.Keywords
Funding Information
- German Research Foundation (DFG) as part of the Transregional Collaborative Research Center “Invasive Computing” (SFB/TR 89)
- National Science Foundation (1329780, 1239246, 1525609)
This publication has 11 references indexed in Scilit:
- Analyzing OpenCL 2.0 workloads using a heterogeneous CPU-GPU simulator2017
- FinePar: Irregularity-aware fine-grained workload partitioning on integrated architectures2017
- Understanding Co-Running Behaviors on Integrated CPU/GPU ArchitecturesIEEE Transactions on Parallel and Distributed Systems, 2016
- A comprehensive performance analysis of HSA and OpenCL 2.02016
- A black-box approach to energy-aware scheduling on integrated CPU-GPU systems2016
- Achieving performance with OpenCL 2.0 on Intel® processor graphics2015
- Adaptive heterogeneous scheduling for integrated GPUs2014
- Fluidic Kernels2014
- gem5-gpu: A Heterogeneous CPU-GPU SimulatorIEEE Computer Architecture Letters, 2014
- CPU-assisted GPGPU on fused CPU-GPU architectures2012