Search for: All records

The Dragonfly networks have been adopted in the current supercomputers, and will be deployed in future generation supercomputers and data centers. Effective routing on Dragonfly is challenging. Universal Globally Adaptive Load-balanced routing (UGAL) is the state-of-the-art routing algorithm for Dragonfly. For each packet, UGAL selects either a minimal path or a non-minimal path based on their estimated latencies. Practical UGAL makes routing decisions with local information, deriving the estimated latency for each path from the local queue occupancy and path hop count information. In this work, we develop techniques to improve the accuracy of the latency estimation for UGAL with local information, which results in more effective routing decisions. In particular, our schemes are able to proactively mitigate the potential network congestion with imbalanced network traffic. Extensive simulation experiments using synthetic traffic patterns and application workloads demonstrate that our enhanced UGAL schemes significantly improve the routing performance for many common traffic conditions. 

he Universal Globally Adaptive Load-balance Routing (UGAL) with global information, referred as UGAL-G, represents an ideal form of adaptive routing on Dragonfly. UGAL-G is impractical to implement, however, since the global information cannot be maintained accurately. Practical adaptive routing schemes, such as UGAL with local information (UGAL-L), performs noticeably worse than UGAL-G. In this work, we investigate a machine learning approach for routing on Dragonfly. Specifically, we develop a machine learning-based routing scheme, called UGAL-ML, that is capable of making routing decisions like UGAL-G based only on the information local to each router. Our preliminary evaluation indicates that UGAL-ML can achieve comparable performance to UGAL-G for some traffic patterns.