Topology-Aware Communication in Wide-Area Message-Passing

Abstract

This position paper examines the use of topology-aware communication services to support message-passing in wide-area, distributed environments, i.e., grids. Grid computing promises great benefits in the flexible sharing of resources but poses equally great challenges for high-performance computing, that is to say, how to execute large-scale computations on a grid with reasonable utilization of the machines involved. For wide-area computations using a message-passing paradigm, these issues can be addressed by using topology-aware communication, i.e., communication services that are aware of and can exploit the topology of the network connecting all relevant machines. Such services can include augmented communication semantics (e.g., filtering), collective operations, content-based and policy-based routing, and managing communication scope to manage feasibility. While such services can be implemented and deployed in a variety of ways, we propose the use of a peer-to-peer, middleware forwarding and routing layer. In a related application domain (time management in distributed simulations) we provide emulation results showing that such topology-awareness can play a major role in performance and scalability. Besides these benefits, such communication services raise a host of implementation and integration issues for their operational deployment and use in grid environments. Hence, we discuss the need for proper APIs and high-level models.