LFF:一种面向大数据应用的众核处理器访存公平性调度机制

研究了众核处理器的访存公平性问题。针对众核处理器距离访存资源较近的处理单元拥有较大的访存带宽而造成的访存公平性问题,提出了一种面向大数据应用的众核处理器访存公平性调度机制:最少最远(LFF)优先访存。这种机制的原理如下:依据处理单元距离访存资源的距离以及处理单元访存的次数来调度访存顺序,以保证各个处理单元的公平性。首先,访问次数较少的节点被赋予更高的访存优先权。其次,在具有相同访问次数的节点中,距离更远的节点优先访存。再次,在相同距离的节点中,已被选中优先次数少的有优先级。实验评估表明,该调度机制能够有效解决众核处理器的访存公平性问题,其公平性调度效果优于FR-FCFS,PAR-BS、ATLAS。在1024核情况下,系统异步率由FR-FCFS的15.5%降低到1.89%。

LFF:A many-core processor's access fairness scheduling scheme for big data applications

The memory access fairness problem of many-core processors was studied.Aiming at the memory access fairness problem that many-core processor's process units closer to the memory controller have higher memory access bandwidth, this study proposed a many-core processor's memory access fairness scheduling scheme for big data applications, called the least and furthest first (LFF) access scheme.This scheme schedules the order of memory access requests according to the distance from a processing unit to its access resource and a processing unit's access history.Firstly, the highest memory access priority is assigned to the nodes with least access request times.Secondly, for the nodes with same access request times, the furthest nodes access memory preferentially.Thirdly, among the nodes with the same distance, the nodes with less priority-assigned history access memory firstly.Our evaluation shows that the proposed scheme can efficiently solve the memory access fairness problem in many-core processors, and its effectiveness in fairness scheduling outperforms the schemes of FR-FCFS, PAR-BS and ATLAS.In the case of 1024 cores,the execution asynchronous rate was reduced to 1.89% from 15.5% compared with the FR-FCFS.