An Address Cache of Interconnect Network in Parallel Computers

Most of users are accustomed to utilize the virtual address in their parallel programs running at the scalable parallel computer systems. Therefore the virtual and physical address translation directly affects the performance of the parallel programs. Cache can strongly improve the efficiency of address translation and reduce the latency of translation. In this paper, a new address translation cache (ATC) is proposed for the interconnect network of scalable parallel computer systems. To improve the hit ratio, ATC adopts embedded dynamic random access memory (eDRAM) to store more address translation table items. A new eDRAM refresh mechanism is proposed to hide the refresh operation and avoid the performance loss introduced by refresh. In ATC, there are many reliability techniques, including error correcting code and a novel bypass module. The well-known NPB benchmarks have been run at the 32 compute nodes including ATC. The results show that the ATC has high hit ratio which the average value of 32 nodes is 95.3%. It is indicated that ATC is well designed and has high performance. It also has been compared with three types of typical cache implemented by different capacities SRAM (static random access memory), and the conclusion is the capacity of cache is key factor to improve the hit ratio.