利用后溯重调度过程优化簇间数据交互

Optimizing inter-cluster data communications through optimized backward rescheduling

在多簇处理器情况下,指令应用所带来的簇间数据交互问题已经成为制约处理器性能的关键问题。针对此问题提出了在一般的调度后进行一次后溯重调度优化过程,减少了簇间的数据交互量,提高了编译器关于处理器的利用率,同时减少了编译生成的指令序列运行时所消耗的功耗。实验结果表明,利用该方法进行调度,比列表调度算法簇间数据交互量减少平均44.36%,调度后的指令执行时间的平均减少量为24.93%,比UAS(unified assign and schedule)调度算法簇间数据交互量减少平均31.25%,调度后的指令执行时间的平均减少量为14.62%。

As clustering becoming more essential for embedded computer systems, compilers for clustered systems must carefully assign instructions in the clusters to minimize the inter-cluster data communications which degrade performance and increase power consumption. This work proposed an optimized backward rescheduling process which runs after the traditional scheduling process. The tests show the rescheduling reduces the inter-cluster communications by 44% on average when compared to list scheduling, and reduces the final execution time by 25%. When compared with UAS （unified assign and schedule） algorithm, the inter-cluster communications are reduced by 31% and the execution time by 15%.