多核共享缓存bank冲突分析及其延迟最小化

Analyzing Bank Access Conflict and Minimizing Bank Conflict Delay for Shared Cache in Multicore

在硬实时多核系统中,共享资源冲突的问题为硬实时任务的最差情况下执行时间(WCET)分析带来了新挑战.虽然现有的共享缓存冲突分析技术在storage冲突方面已取得研究进展,但对于bank冲突而言,现有研究仍局限于通过界定bank冲突延迟上限来分析和处理bank冲突.该文通过优化核-bank映射关系来使硬实时多核系统中的bank冲突延迟最小化,即在对bank冲突延迟进行分析的基础上,首先通过优化核-bank之间的映射关系来消除bank冲突;若无法消除,则需要寻找能使bank冲突延迟最小化的核-bank映射关系解,并为此设计了一种基于多核总线请求时间序列的bank冲突延迟求解算法.最后,文中设计了能够对总线访问延迟进行消重的多核硬实时任务WCET估算方法.实验结果表明:文中所提的优化方法可消除这类bank冲突或使其延迟最小化,文中所提的WCET估算方法与现有估算方法相比可获得更精确的最差情况下执行时间(WCET).

Inter-task interferences on the shared resources of hard real-time multicore systems bring a new challenge to WCET analysis. The inter-task interferences related to the shared cache are storage interference and bank access conflict. Up to now, storage interference has been solved well. In existing research, however, the treatment of bank access conflict is only confined to bounding upper-bound of the bank conflict delay suffered by one request. As changing the core- to-bank mapping can change the bank access conflict in a hard real-time multicore system using a bank partitioned shared cache, we optimize core-to-bank mapping to minimize bank conflict delay in this paper. We firstly optimize core-to-bank mapping to eliminate bank access conflict. If cannot eliminate the bank access conflict, we optimize core-t〇 -bank mapping to minimize bank conflict delay. To solve the optimization problem, we design an algorithm using the timing sequences of the bus requests to compute the bank conflict delay on one shared bank. We also design a method to estimate the WCETs of hard real-time tasks in multicore systems, which can reduce the time overlapping among the execution time in pipeline, bus access delay and the latency of memory system. Experimental results demonstrate that our approach of optimizing core-to-bank mapping can minimize bank conflict delay and our approaches to estimate WCET are more effective than existing approaches.