Xin Zhao, Pavan Balaji, William Gropp, and Rajeev Thakur won the Best Paper Award at the 13th IEEE International Conference on Scalable Computing and Communication (ScalCom2013) for their paper “Optimization Strategies for MPI-Interoperable Active Messages.”
Data-intensive applications, such as those in bioinformatics and social network analysis, often involve irregular computation and communication patterns, making them ill-suited for traditional data movement approaches. Zhao, a Ph.D. student at the University of Illinois in Urbana-Champaign, worked at Argonne as a research aide enhancing an alternative programming model, called active messages (AMs), that enables researchers to dynamically move a computation closer to the data, rather than moving the data to the local process.
Previous work on AMs typically has involved low-level, machine-specific libraries that do not support the popular Message Passing Interface (MPI) or has focused on implementations that support AMs on top of the MPI library but that lack capabilities such as memory consistency.
To address these limitations, Zhao and her colleagues developed optimization strategies designed to improve the performance of AMs. In addition, the team conducted a detailed evaluation of their optimization strategies on a 4096-core cluster. Their results demonstrate considerable performance advantages from the new strategies.
For example, the paper describes a way to reduce redundant synchronization, either automatically or through “user hints.” In tests of the benefit of hints under increasing network contention, the researchers were able to show that even when a given hint required coordination of active messages with the target, the performance improved by 50% at scale.
ScalCom2013 was held Dec. 21-22, 2013 in Chengdu, China, in conjunction with the IEEE International World-Ubiquitous Science Congress. The aim of ScalCom 2013 was to address research challenges and present advances in the design and implementation of scalable computing and communication systems.