Best Paper Award
The Case for Network Coding for Collective Communication on HPC Interconnection Networks
Ahmed SHALABY，Ikki FUJIWARA，Michihiro KOIBUCHI
[Trans. Inf. & Syst., Mar. 2015]

Ahmed SHALABY
Ikki FUJIWARA
Michihiro KOIBUCHI		  
This paper proposed a network coding scheme which can reduce the number of unicasts and the transfer data size in collective communication on HPC interconnection networks. Network coding techniques can optimize data flow to improve the throughput and efficiency of the network. The proposed network coding scheme has a hierarchical multicasting structure with intra-group and inter-group unicasts and can reduce the number of unicasts and the size of the data transferred in latency-sensitive collective communications.
This paper shows the effectiveness of the proposed network coding by quantitative analysis and cycle-accurate network simulations.
Quantitative analysis revealed that in a 4,096-switch network, the proposed network coding can decrease the aggregate path hop counts by as much as 94% compared to conventional unicast-based multicasts. The authors also evaluate the performance of the hierarchical network coding using a cycle-accurate network simulator called BookSim. The results show that the proposed network coding invariably improves the execution time. They also demonstrate that as the network size and the packet length increase, the benefits become greater. In addition, the authors evaluate hierarchical network coding with different latency overheads at the intermediate nodes. The results show that the overheads at the intermediate nodes only marginally affect the end-to-end latency and it is not a bottleneck for collective communications.
This research resulted in an improvement in the performance of network coding and made a significant contribution to development of the HPC field.
This paper also contains a detailed discussion on parameter optimization based on quantitative analysis and cycle-accurate simulations.

Close