论文部分内容阅读
在基于微片(flit)分组的动态缓存分配基础上,提出一种基于微片分组的片上网络交叉开关调度机制.该机制与静态独立分割缓存的思想不同,首先对输入端缓存进行统一管理,对微片根据其流向进行分组,并为所分各“组”动态分配缓存,然后引入一种基于“组”规模的概率仲裁算法,通过“组”分配和开关分配实现调度过程.为进一步降低开销,还在该机制基础上提出一种各“组”共享仲裁的策略.理论分析与实验结果均表明:所提出的机制相对于传统和动态虚通道机制,可节约25%以上的硬件开销并可获得更优的网络延迟与吞吐性能;共享仲裁策略可在所提机制基础上进一步降低硬件开销,但其代价是网络性能有所下降.
Based on the dynamic cache allocation based on flit packet, this paper proposes a chip-based on-chip crossbar switch scheduling mechanism based on microchip grouping. This mechanism is different from the idea of static independent partitioning cache. First, the input buffer is managed uniformly, The flitches are grouped according to their flow direction, and the caches are dynamically allocated to the subgroups. Then a probabilistic arbitration algorithm based on the “group” scale is introduced, which is implemented by “group ” allocation and switch assignment Scheduling process.In order to further reduce the overhead, a strategy of sharing arbitration is proposed on the basis of this mechanism.The theoretical analysis and experimental results show that the proposed mechanism can be compared with the traditional and dynamic virtual channel mechanism Save more than 25% of the hardware overhead and obtain better network latency and throughput performance. The shared arbitration strategy can further reduce the hardware overhead based on the proposed mechanism, but at the cost of reduced network performance.