超混合深度可重构计算阵列调度策略的优化研究

STUDY ON SCHEDULING POLICY OPTIMISATION BASED ON HETEROGENEOUS RECONFIGURABLE COMPUTING ARRAY

针对一种新型的高性能计算机结构:超混合深度可重构计算机阵列(HRCA),提出两个在HRCA上任务分配的调度优化方法。(1)通过算核的优化分配减轻或消除由于算核分配引起的数据通信量急剧增加而导致的"存储墙"问题;(2)通过算粒的调度,将两次迭代间的数据交换与计算时间相重叠,缩短计算部件由数据交换导致的等待时间。以N-body FMM算法为例,验证了两种方法有效地降低了系统对于片外存储访问速度需求,提高了系统的利用率。

Aiming at a novel high performance computer architecture, the heterogeneous reconfigurable computing array （ HRCA ）, we propose two optimisation methods for the scheduling of computing tasks allocation on HRCA. （ 1 ） Relieving or eliminating by optimised computing cores allocation the ＂memory wall＂ problem caused by the surge of data communication amount due to computing cores allocation; （2） Overlapping the time of computing and data exchanging between two iterations by scheduling task kernels to shorten the idle time on computing components due to data exchange. We use the N-body FMM algorithm as the example and have verified that these two methods effectively reduce the access speed requirements of the system on off-chip memory and improve the utilisation of the system.