Contention resolution mechanisms for asynchronous optical packet switching based high performance computing system

Authors:
Jun Zhao;Xiaohan Sun
Affiliations:
Lab of Photonics and Optical Communications, Dept of Electronics Engineering, Southeast University, Nanjing, China 210096;Lab of Photonics and Optical Communications, Dept of Electronics Engineering, Southeast University, Nanjing, China 210096
Venue:
The Journal of Supercomputing
Year:
2010

Citing 3
Cited 0

Challenges in Building a Flat-Bandwidth Memory Hierarchy for a Large-Scale Computer with Proximity Communication

HOTI '05 Proceedings of the 13th Symposium on High Performance Interconnects
Designing a Crossbar Scheduler for HPC Applications

IEEE Micro
The Data Vortex, an All Optical Path Multicomputer Interconnection Network

IEEE Transactions on Parallel and Distributed Systems

Quantified Score

Hi-index	0.01

Visualization

Abstract

We propose an asynchronous optical packet switching (AOPS) based multistage interconnection network employing the distributed control and banyan-like routing scheme, which can construct the high performance computing system (HPCS). A flexible, low latency, and high capacity recycling-fiber-delay-line (Rec-FDL) based collision resolution mechanism is proposed with higher switching efficiency for burst services than that by the conventional packet retransmission algorithm (C-PRA). The system performances, including stability, blocking rate (BR), and longest packets waiting latency (LPWL) are analyzed. The simulation shows that the system stability condition for the Rec-FDL mechanism is the new packet arriving rate 驴 less than 30脳106 packets/s. In addition, the BR and Rec-FDL induced LPWL are 0.1235---0.1413 and 6.305 ns---7.651 ns, respectively, with increasing the number of CPUs connected to the system from 4 to 16,384, while the increment for LPWL induced by the conventional algorithm is 3.9 ns at the same conditions as above. Therefore, the mechanism presented here is more beneficial for the system scalability and has higher latency reducing abilities than the C-PRA.