Abstract

We present a system demonstration of a sequential deadlock recovery-based scheme [1] employing a circulating token to guarantee mutual exclusive access to deadlock buffers. This system allows us to evaluate the design and implementation of a multiprocessor network router using free-space optical interconnects. The system is designed based on our CMOS-SEED chip—OMNI [2] and GaAs-based chip—WARRP [3]. These two chips are employed to implement asynchronous optical token-based resource arbitration and deadlock recovery via channel preemption, respectively. With an external node controller, this experimental system is capable of demonstrating a high-performance bandwidth-efficient deadlock recovery-based multiprocessor network router.

© 1997 Optical Society of America