Argonne National Laboratory

Dynamic Time-Variant Connection Management for PGAS Models on InfiniBand

TitleDynamic Time-Variant Connection Management for PGAS Models on InfiniBand
Publication TypeConference Paper
Year of Publication2013
AuthorsVishnu, A, Krishnan, M, Balaji, P
Conference NameThe Third International Workshop on Accelerators and Hybrid Exascale System, IEEE International Parallel and Distributed Processing Symposium
Conference LocationBoston, MA
Other NumbersANL/MCS-P5044-1213

InfiniBand (IB) has established itself as a promising network infrastructure for high-end cluster computing systems as evidenced by its usage in the Top500 supercomputers today. While the IB standard describes multiple communication models (including reliable-connection (RC), and unreliable datagram (UD)), most of its promising features such as remote direct memory access (RDMA), hardware atomics and network fault tolerance are only available for the RC model which requires connections between communicating process pairs. In the past, several researchers have proposed on-demand connection man- agement techniques that establish connections when there is a need to communicate with another process. While such tech- niques work well for algorithms and applications that only communicate with a small set of processes in their life-time, there exists a broad set of applications that do not follow this trend. For example, applications that perform dynamic load balancing and adaptive work stealing have a small set of communicating neighbors at any given time, but over time the total number of neighbors can be very high; in some cases, equal to the entire system size.

In this paper, we present a dynamic time-variant connection management approach that establishes connections on-demand like previous approaches, but further intelligently tears down some of the unused connections as well. While connection tear- down itself is relevant for any programming model, different models have different complexities. In this paper, we study the Global Arrays (GA) PGAS model for two reasons: (1) the simple one-sided communication primitives provided by GA and other PGAS models ensure that connection requests are always initiated by the origin process without explicit synchronization with the target process—this makes connection tear-down simpler to handle; and (2) GA supports applications in several domains such as computational chemistry (NWChem) and computational biology (ScalaBLAST) that demonstrate this behavior making it an obvious first target for the proposed enhancements. We evaluate our proposed approach using NWChem computational chemistry application using up to 6144 processes, and show that our approach can significantly reduce the memory requirements of the communication library while maintaining its performance.