基于重用距离分析的多虚拟机缓存划分系统

发布时间：2019-01-19 10:52

【摘要】：随着多核和众核技术的普及，计算机以并行化计算方式大幅度提升了系统整体的处理能力和计算性能。但是多核技术也引入了严重的资源竞争问题，各并发任务共享缓存等硬件，造成了缓存资源竞争，导致竞争应用缓存命中率降低，产生严重的性能损失。尤其是针对高性能计算的虚拟化集群环境，缓存资源竞争导致虚拟机之间的隔离机制被打破，带来计算性能损失和安全可靠性降低的双重问题。基于重用距离分析的多虚拟机缓存划分系统，通过对高性能计算应用进行访存地址采样和缓存缺失率在数据集上变动情况的性能采样，使用分割碎片计算，循环压缩数据方法加快大规模采样数据的处理，生成各应用的重用距离直方图，记录各应用的重用距离在缓存块上的分布比例情况。对重用距离等缓存行为数据的分析过程，使用特征类型识别技术，获得各应用的缓存行为特征类型。虚拟机缓存划分机制接收用户的特化应用虚拟资源申请，对各虚拟机的应用进行特征类型匹配，重点处理缓存污染型和缓存敏感型应用虚拟机的资源分配；通过节点间的通信协作，，完成各虚拟机的资源创建和物理核资源的隔离绑定，实现虚拟机缓存划分。系统通过应用的重用距离分析识别应用类型，在虚拟机隔离的粒度上将各应用间缓存影响降到最低，实现虚拟机应用间的缓存划分隔离效果，提升了高性能计算应用的计算性能和服务质量。测试结果分析表明，采用基于重用距离分析的多虚拟机缓存划分系统，对缓存敏感型应用有良好的性能提升效果，包括运行时间缩短10%~21%，吞吐量增加10%~23%，缓存缺失率降低10%~25%。
[Abstract]:With the popularization of multi-core and multi-kernel, computer greatly improves the whole processing power and computing performance of the system by parallel computing. However, the multi-core technology also introduces the serious problem of resource competition. The concurrent tasks share cache and other hardware, which results in the cache resource competition, which leads to the decrease of cache hit ratio and the serious performance loss in competing applications. Especially for the virtualized cluster environment of high performance computing, the contention of cache resources leads to the breaking of the isolation mechanism between virtual machines, which leads to the loss of computing performance and the decrease of security and reliability. A multi-virtual machine cache partition system based on reuse distance analysis is proposed. By sampling the memory access address of the high performance computing application and the performance sampling of the change of cache missing rate on the data set, the partition fragment calculation is used. The cyclic compression method speeds up the processing of large-scale sampled data, generates the reuse distance histogram of each application, and records the distribution ratio of the reuse distance of each application on the buffer block. In the analysis process of cache behavior data such as reuse distance, the feature type recognition technique is used to obtain the cache behavior feature types of each application. The virtual machine cache partition mechanism receives the user's application virtual resource request, carries on the characteristic type matching to each virtual machine application, and deals with the resource allocation of the cache polluted and cache sensitive application virtual machine. Through the communication and cooperation between nodes, the resource creation of each virtual machine and the isolation binding of the physical core resources are completed, and the cache partition of the virtual machine is realized. By analyzing the reuse distance of the application to identify the application type, the system minimizes the impact of cache between the applications on the granularity of the virtual machine isolation, and realizes the partition and isolation effect of the cache between the virtual machine applications. The computing performance and service quality of high performance computing applications are improved. The test results show that the multi-virtual machine cache partition system based on reuse distance analysis can improve the performance of the cache sensitive application, including shortening the running time by 10 / 21 / 21, increasing the throughput by 10 / 23, and improving the performance of the system. The cache missing rate is reduced by 10% and 25%.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP302

【参考文献】