Routing algorithms for the emerging topologies in HPC and data center networks.

摘要

The computing capabilities of the high performance computing systems have increased manyfold over the span of years following Moores law and have achieved petascale recently. Following the trend, the computation cababilities are expected to reach exascale by the end of this decade. The strive for exascale computing poses significant challenge on the processor system design as well as the design of network architectures. Several network optimizations pertaining to topology and routing have been devised for the existing architectures and new architectures are also being proposed to meet the design goals of the next generation supercomputers. As an important architectural component, new topologies are being proposed and coming into existence. Topologies exhibiting excellent properties might suffer from sub-optimal performance if the routing schemes supported on them are not able to utilize the full potential. An effective routing scheme that fully utilizes the network resources of a topology is the key to meeting some of the design goals of network architectures. The goal of my research is to investigate the emerging topologies and provide optimized routing schemes that utilize the network efficiently with an aim to provide high throughput and low latency communications. In particular, I consider fat-trees and their slimmed versions, and the recently proposed jellyfish topology and propose effective routing schemes on these topologies. The developed schemes show improvements over the existing routing schemes.