基于组件的应用中的动态GPU资源管理
Dynamic GPU Resource Management in Component-Based Applications
DOI: 10.12677/CSA.2023.135093, PDF,   
作者: 罗 婧, 魏 雄：武汉纺织大学，计算机与人工智能学院，湖北 武汉
关键词: GPU；资源管理；空间多任务处理；优化；GPU； Resource Management； Spatial Multitasking； Optimization
摘要: 在边缘和云环境中，使用图形处理单元(GPU)作为高速并行计算设备可以提高计算密集型应用程序的性能。随着要处理的数据量和复杂性的增加，多个相互依赖的组件序列在GPU上共存并共享GPU资源。由于缺乏用于动态GPU资源分配的低开销和在线技术会导致GPU使用不平衡并影响整体性能，提出了高效的GPU内存和资源管理器。管理器通过使用共享内存和动态分配部分共享GPU资源来提高整体系统性能。评估结果表明，与默认GPU并发多任务处理相比，动态资源分配方法能够将具有各种并发组件数的应用程序的平均性能提高29.81%。同时，使用共享内存可使性能提高2倍。
Abstract: In edge and cloud environments, using graphics processing units (GPUs) as high-speed parallel computing devices can improve the performance of computing intensive applications. As the amount and complexity of data to be processed increases, multiple interdependent component sequences coexist on the GPU and share GPU resources. Due to the lack of low overhead and online technology for dynamic GPU resource allocation, which can lead to uneven GPU usage and affect overall performance, an efficient GPU memory and resource manager is proposed. The manager improves overall system performance by using shared memory and dynamically allocating partial shared GPU resources. The evaluation results indicate that the dynamic resource allocation method can improve the average performance of applications with various concurrent component counts by 29.81% compared to the default GPU concurrent multitasking processing. At the same time, using shared memory can increase performance by 2×.
文章引用：罗婧, 魏雄. 基于组件的应用中的动态GPU资源管理[J]. 计算机科学与应用, 2023, 13(5): 953-963. https://doi.org/10.12677/CSA.2023.135093

参考文献

[1] Capodieci, N., Cavicchioli, R., Bertogna, M., et al. (2018) Deadline-Based Scheduling for GPU with Preemption Support. 2018 IEEE Real-Time Systems Symposium (RTSS), Nashville, TN, USA, 11-14 December 2018. [Google Scholar] [CrossRef] 
[2] Adriaens, J.T., Compton, K., Kim, N.S., et al. (2012) The Case for GPGPU Spatial Multitasking. IEEE International Symposium on High Performance Computer Architecture, New Orle-ans, LA, USA, 25-29 February 2012. [Google Scholar] [CrossRef] 
[3] Nardin, I., Righi, R., Lopes, T., et al. (2021) On Revisiting En-ergy and Performance in Microservices Applications: A Cloud Elasticity-Driven Approach. Parallel Computing, 108, Article ID: 102858. [Google Scholar] [CrossRef] 
[4] Chen, L., Zigerelli, A., Yang, J., et al. (2018) A Dynamic and Proactive GPU Preemption Mechanism Using Checkpointing. IEEE Transactions on Computer-Aided Design of Inte-grated Circuits and Systems, 39, 75-87. [Google Scholar] [CrossRef] 
[5] Garg, S., Kothapalli, K. and Purini, S. (2018) Share-a-GPU: Providing Simple and Effective Time-Sharing on GPUs. 2018 IEEE 25th International Conference on High Performance Computing (HiPC), Bengaluru, India, 17-20 December 2018. [Google Scholar] [CrossRef] 
[6] Zhao, C., Gao, W., Nie, F., et al. (2022) A Survey of GPU Multitasking Methods Supported by Hardware Architecture. IEEE Transactions on Parallel and Distributed Systems, 33, 1451-1463. [Google Scholar] [CrossRef] 
[7] Liang, Y., Huynh, H.P., Rupnow, K., et al. (2015) Efficient GPU Spatial-Temporal Multitasking. IEEE Transactions on Parallel and Distributed Systems, 26, 748-760. [Google Scholar] [CrossRef] 
[8] NVIDIA (2022) Multi-Process Service.
https://docs.nvidia.com/deploy/pdf/CUDA_Multi_Process_Service_Overview.pdf
[9] Aguilera, P., Lee, J., Farmahini-Farahani, A., et al. (2014) Process Variation-Aware Workload Partitioning Algorithms for GPUs Supporting Spatial-Multitasking. 2014 Design, Automation & Test in Europe Conference & Exhibition (DATE), Dresden, Germany, 24-28 March 2014, 726-731. [Google Scholar] [CrossRef] 
[10] Aguilera, P., Morrow, K. and Kim, N.S. (2014) Fair Share: Allocation of GPU Resources for Both Performance and Fairness. 2014 IEEE 32nd International Conference on Computer Design (ICCD), Seoul, South Korea, 19-22 October 2014. [Google Scholar] [CrossRef] 
[11] Zhang W, Chen Q, Zheng N, et al. (2021) Towards QoS-Awareness and Improved Utilization of Spatial Multitasking GPUs. IEEE Transactions on Computers, 71, 866-879. [Google Scholar] [CrossRef]