SIParCS 2017-Ananta Thapaliya
Supercomputing Infiniband Fabric Analysis
InfiniBand is a powerful network-communication standard that provides fast interconnections between various end nodes and storage systems in supercomputers. Monitoring network performance helps ensure efficiency in high performance computing environments, but supporting technologies also add significant value. Static analysis of the network architecture is crucial to maintaining an efficient data-transfer network. This research focuses on visual analysis of the InfiniBand architecture using a graphical representation of the switches and end-nodes as nodes, with the cables as edges. This visual representation results from parsing the information rendered by the ‘ibnetdiscover’ command in NCAR’s ‘InfiniBand Import’ plugins for Tulip, an extensible open source software for graphical visualization. We used C++ to develop Tulip Plugins that employ network analysis algorithms on the graph, then we compared the route suggested by those algorithms against the physical topology. All the plugins were built, installed, and tested using a software container technology called ‘Docker.’ The final products of this project were the Tulip Plugins and the Docker container that runs them on Tulip.
In average, the logical routes proved to be less optimal than the ideal route suggested by Dijkstra’s algorithm. This leads to the conclusion that the data interconnect as it is works efficiently but could be improved by twisting the subnet manager a little bit. In high performance environments like supercomputers, data packets taking fewer hops substantially betters the overall performance, especially while several jobs are being run at the same time.
Mentors: Nathan Rini, Tom Kleespies