Abstract

As feature sizes decrease to the submicrometer regime and clock rates increase to the multigigahertz range, the limited bandwidth at higher bit rates and longer communication distances in electrical interconnects will create a major bandwidth imbalance in future high-performance computing (HPC) systems. We explore the application of an optoelectronic interconnect for the design of flexible, high-bandwidth, reconfigurable and adaptive interconnection architectures for chip-to-chip and board-to-board HPC systems. Reconfigurability is realized by interconnecting arrays of optical transmitters, and adaptivity is implemented by a dynamic bandwidth reallocation (DBR) technique that balances the load on each communication channel. We evaluate a DBR technique, the lockstep (LS) protocol, that monitors traffic intensities, reallocates bandwidth, and adapts to changes in communication patterns. We incorporate this DBR technique into a detailed discrete-event network simulator to evaluate the performance for uniform, nonuniform, and permutation communication patterns. Simulation results indicate that, without reconfiguration techniques being applied, optical based system architecture shows better performance than electrical interconnects for uniform and nonuniform patterns; with reconfiguration techniques being applied, the dynamically reconfigurable optoelectronic interconnect provides much better performance for all communication patterns. Based on the performance study, the reconfigured architecture shows $30\%–50\%$ increased throughput and $50\%–75\%$ reduced network latency compared with HPC electrical networks.

© 2009 Optical Society of America

RAPID for high-performance computing systems: architecture and performance evaluation

Avinash Karanth Kodi and Ahmed Louri
Appl. Opt. 45(25) 6326-6334 (2006)

Hierarchical and reconfigurable optical/electrical interconnection network for high-performance computing

Zuoqing Zhao, Bingli Guo, Yu Shang, and Shanguo Huang
J. Opt. Commun. Netw. 12(3) 50-61 (2020)

Performance trade-offs in reconfigurable networks for HPC

Min Yee Teh, Zhenguo Wu, Madeleine Glick, Sebastien Rumley, Manya Ghobadi, and Keren Bergman
J. Opt. Commun. Netw. 14(6) 454-468 (2022)

SPRINT: Scalable Photonic Switching Fabric for High-Performance Computing (HPC)

Brian Neel, Randy Morris, Dominic Ditomaso, and Avinash Kodi
J. Opt. Commun. Netw. 4(9) A38-A47 (2012)

Optical High-Performance Computing—JOSA A and Applied Optics: Introduction

H. John Caulfield, Shlomi Dolev, and William M. J. Green
J. Opt. Soc. Am. A 26(8) OHPC1-OHPC2 (2009)