Topology Aware Routing Techniques for Next Generation Interconnect Networks

As the world is moving towards exascale computing, interconnect networks are becoming more and more important because of their omnipresent use in high performance computing systems and in large scale data centers. The performance of an interconnect network depends on its topology, routing, job distribution, and other technological factors, and can become a major performance bottleneck for the entire system. My research as a PhD candidate in Computer Science in Florida State University is focused on interconnect network architecture. To be precise, I work to design topology-aware adaptive routing schemes for some existing and proposed interconnect topologies which will improve the performance of the respective systems. First, I perform a comprehensive analysis on Slim Fly network topology and demonstrate that the topology in its original form is not load-balanced. Because of the way the topology is formed, certain links are more likely to be utilized than the rest of the links, which leads the system to perform in less than optimum level. I propose two novel schemes to address the issue. The first scheme involves modifying the topology by allocating more band-width to the heavily-used links. The second approach modifies the adaptive routing used over the topology to redistribute the traffic flows to achieve better load balance. Second, I notice that the fraction of shorter minimal and non-minimal paths can vary across the design space of the Dragonfly topology, but traditional adaptive routing does not take advantage of this. I propose Topology-Custom UGAL routing (T-UGAL) for Dragonfly that customizes the set of the non-minimal paths used in UGAL in accordance to the topology underneath, which leads to shorter average path lengths and better system performance in terms of packet latency and system throughput. I design a multi-step algorithm to find the most optimum set of non-minimal paths for the particular topology. For both of my projects, I follow the common theme of discovering inherent properties of the topology, and modifying the routing scheme to leverage those characteristics. Considering the fact that large HPC systems often need significant investments to construct, and tend to retain their structure over years, this is a robust and practical approach to ensure the fullest utilization of the systems.