Modeling Network Contention Effects on All-to-All Operations

Abstract

One of the most important collective communication patterns used in scientific applications is the complete exchange, also called All-to-All. Although efficient complete exchange algorithms have been studied for specific networks, general solutions like those available in well-known MPI distributions (e.g. the MPI_Alltoall operation) are strongly influenced by the congestion of network resources. In this paper we present an integrated approach to model the performance of the All-to-All collective operation. Our approach consists in identifying a contention signature that characterizes a given network environment, using it to augment a contention-free communication model. This approach allows an accurate prediction of the performance of the All-to-All operation over different network architectures with a small overhead. This approach is assessed by experimental results using three different network architectures, namely Fast Ethernet, Gigabit Ethernet and Myrinet.

Efficacy Analysis of a SDN-enhanced Resource Management System through NAS Parallel Benchmarks

“…One of such approaches, namely, optimization of process allocation [20]- [22] has been raised. In [20], to optimize MPI collective communications on multiple clusters connected by wide-area networks, the parameterized Log P model, which is a hierarchical extension of the Log P model, has been proposed.…”

“…In [20], to optimize MPI collective communications on multiple clusters connected by wide-area networks, the parameterized Log P model, which is a hierarchical extension of the Log P model, has been proposed. In [21] and [22], by taking the collision of communications generated by MPI processes into consideration, degradation of communication performance is suppressed. These approaches enable an application to efficiently communicate between MPI processes by taking the topology of the interconnect and the communication pattern of the application into consideration.…”

Extending τ-Lop to model concurrent MPI communications in multicore clusters

“…LoGPC [14] introduces a large set of parameters to evaluate contention in mesh networks with wormhole routing. Work [39] introduces the contention factor, represented as a parameter exclusively affecting the network performance, which is incremented linearly from the content-free communication cost. The former work evaluates the total exchange collective under this assumption.…”

Adaptive approaches for efficient parallel algorithms on cluster-based systems

“…There are many parallel communication models in the literature that analyze performances based on system parameters (see Alexandrov et al (1995); Culler et al (1996); Faik et al (2005); Frank et al (1997); Hockney (1994); Kielmann et al (2001); Lastovetsky et al (2006); Moritz and Frank (2001); Teresco et al (2005)). More sophisticated models have been proposed for complex architectures (Cappello et al (2005); Steffenel (2006)). All these models differ on the assumptions about the computational support parameters, such as latency, heterogeneity, network con-tention, etc., and therefore are able to cover a great variety of architectures and modeling aspects.…”