Deterministic and chaotic adaptive routing in multicomputers

Multicomputers such as the Caltech Cosmic Cube and its commercial descendants, employing distributed memory and message passing, manage to avoid the scalability problems of shared-memory multiprocessors, thus becoming a promising approach to massive parallelism. The efficient routing of messages is one of the fundamental issues in multicomputers. State-of-the-art systems implement oblivious routing, where the path of each message is unique and completely determined by its source and destination addresses. Oblivious routers, although simple to implement, exhibit poor performance under traffic known to create hot-spots and they are fault-intolerant. Adaptive routers, an alternative to obliviousness, select message paths based on the local load characteristics. As a result they are better at diffusing congestion and they are more fault-tolerant. In this thesis, two novel adaptive routing schemes are introduced, a minimal and nonminimal one. In the minimal router, the Zenith router, messages are always routed closer to their destination and message delivery can be deterministically guaranteed. In the nonminimal router, the Chaos router, messages can be occasionally sent further from their destination. The hardware requirements and control complexity of the Chaos router are superior to that of any minimal router. This reduction in hardware stems from the fact that the Chaos router does not need to use priority routing, like other nonminimal routers in order to guarantee message delivery. Rather, randomization in the message derouting process is used to guarantee message delivery with high probability. The performance of the suggested routers is evaluated through simulations and their hardware requirements are considered.