Live Webcast 15th Annual Charm++ Workshop

The divergence of application behavior from optimal network usage leads to performance bottlenecks induced by communication. Communication performances are known to worsen when dealing with large quanities of small messages, due to the overhead of envelopes and going through the communication stack multiple times. Prior work has attempted to mitigate this through the aggregration of small messages, but it has only studied the impact for cases where the size of the message is constant and known ahead of time. This thesis explores the applicability of this optimization to variable-sized messages and machines with a large number of cores, analyzing both the theoretical considerations involved and the performance gains achieved in practice. The work is implemented as an update to the Topological Routing and Aggregation Module (TRAM) of the Charm++ parallel programming system.