CPU-GPU异构平台的性能优化研究及其在实时信号模拟技术中的应用

发布时间：2018-06-08 00:48

本文选题：CPU-GPU异构平台 + 并行计算　；参考：《浙江大学》2017年硕士论文

【摘要】：随着GPU的计算能力和可编程性的不断提高,基于CPU-GPU的异构计算逐渐从科学界走向工业界,成为当前的研究热点。相比于同构系统,异构系统具备更优的性能,但同时也带来了更大的程序设计复杂度和性能优化难度。信号模拟器被广泛地运用在探测器和传感器相关工程领域中辅助其系统设计和验证,其回波信号模拟算法需要处理复杂的控制逻辑和大量的数据计算,这与异构系统有着较好的匹配。因此,研究信号模拟在异构平台上的实现技术具有重要的实际应用价值。本文主要展开了以下研究工作:1)整合了 CPU-GPU异构平台上的CPU端和GPU端优化技术,对整个平台进行了全面的性能优化探索。同时,提出了组包和分片数据传输、计算和通信流水化与内核融合等新颖的优化技术,并进行了理论分析和实验验证。2)为了进一步验证各种优化技术的效果,并完整地掌握在CPU-GPU异构平台上进行程序设计的流程,本文以解决复杂电子系统中的回波信号模拟问题为例,对整个系统进行了相关的框架设计、编程实现和性能优化。综合测试结果显示,本文设计的基于CPU-GPU异构平台的实时信号模拟器可达到5.79GB/s的处理能力,相比于同构平台获得了平均21倍的加速比,且可弹性支持扩展多GPU。
[Abstract]:With the improvement of GPU's computing power and programmability, the heterogeneous computing based on CPU-GPU has gradually moved from scientific to industrial, and has become a hot research topic. Compared with the isomorphism system, the heterogeneous system has better performance, but it also brings more complexity of program design and difficulty of performance optimization. Signal simulator is widely used in the field of detector and sensor related engineering to assist its system design and verification. Its echo signal simulation algorithm needs to deal with complex control logic and a large amount of data calculation. This has a good match with the heterogeneous system. Therefore, it has important practical application value to study the realization technology of signal simulation on heterogeneous platform. In this paper, the following research work is carried out: (1) integrating CPU and GPU optimization techniques on CPU-GPU heterogeneous platform, and exploring the overall performance optimization of the whole platform. At the same time, some new optimization techniques, such as packet and slice data transmission, computation and communication pipelining and kernel fusion, are proposed, and theoretical analysis and experimental verification. 2) in order to further verify the effectiveness of various optimization techniques, And master the program design flow on CPU-GPU heterogeneous platform completely. This paper takes solving the echo signal simulation problem in complex electronic system as an example, carries on the related frame design, the programming realization and the performance optimization to the whole system. The test results show that the real-time signal simulator based on CPU-GPU heterogeneous platform can achieve the processing capability of 5.79 GB / s. Compared with the isomorphic platform, the average speedup is 21 times, and the flexibility can support the expansion of multi-GPUs.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP332

【相似文献】