A framework for adaptive collective communications for heterogeneous hierarchical computing systems

Abstract

Collective communication operations are widely used in MPI applications and play an important role in their performance. However, the network heterogeneity inherent to grid environments represent a great challenge to develop efficient high performance computing applications. In this work we propose a generic framework based on communication models and adaptive techniques for dealing with collective communication patterns on grid platforms. Toward this goal, we address the hierarchical organization of the grid, selecting the most efficient communication algorithms at each network level. Our framework is also adaptive to grid load dynamics since it considers transient network characteristics for dividing the nodes into clusters. Our experiments with the broadcast operation on a real-grid setup indicate that an adaptive framework allows significant performance improvements on MPI collective communications.

A Combined Approach for Implementation of Broadcast Operation in Grid Computing

“…The links between pairs of processors are bidirectional, and each process can transmit data on at most one link and receive data on at most one link at any given time. This model is well-known in the literature as 1-port full duplex (Steffenel & Mounié, 2008).…”

“…In a broadcast operation, a so-called root process sends the same message of size m to all other (P-1) processes (Steffenel & Mounié, 2008;Brachet-Estefanel & Mounié, 2006). The simplest way to broadcast a message through all nodes of a grid is to build the spanning tree corresponding to its global topology.…”

An Approximation Algorithm and Dynamic Programming for Reduction in Heterogeneous Environments

“…neither the sender nor the receiver of an ongoing communication can engage in any other communication at the same time. This is referred to as 1-port model in the literature [2,3,22,25,32]. Although communication hardware often allows some degree of overlapping, e.g., a receiver may receive several messages at nearly the same time by using MPI non-blocking routines, this degree of communication parallelism may be limited because processors may only have only limited resource for communication, and multiple communications using the same resource eventually have to be serialized.…”