Characterizing the Communication Demands of the Graph500 Benchmark on a Commodity Cluster

Abstract

. Abstract can be read here. Copyright belongs to IEEE.Abstract-Big Data applications have gained importance over the last few years. Such applications focus on the analysis of huge amounts of unstructured information and present a series of differences with traditional High Performance Computing (HPC) applications. For illustrating such dissimilarities, this paper analyzes the behavior of the most scalable version of the Graph500 benchmark when run on a state-of-the-art commodity cluster facility. Our work shows that this new computation paradigm stresses the interconnection subsystem.In this work, we provide both analytical and empirical characterizations of the Graph500 benchmark, showing that its communication needs bound the achieved performance on a cluster facility. Up to our knowledge, our evaluation is the first to consider the impact of message aggregation on the communication overhead and explore a tradeoff that diminishes benchmark execution time, increasing system performance.

A scalable synthetic traffic model of Graph500 for computer networks analysis

“…Two big levels typically represent around 80 % of the total. The impact of the point‐to‐point messages to notify the end of a tree level is negligible . The likewise sporadic all‐reduce is important only because it synchronizes the generation of new messages.…”

“…Our model consists of a traffic generator structured in multiple stages corresponding to the consecutive tree levels. For each stage, a given number of point‐to‐point messages is dispatched from every process; these messages are uniformly distributed in time and across destinations (following the analysis of the benchmark communications in Fuentes et al) and are succeeded by end‐of‐phase signals and a synchronization point.…”

“…Any given set of nodes can be tuned to under‐ or out‐perform the rest of the system, in order to model heterogenous systems. Message destination is selected randomly among all the other processes in the application, following the uniform distribution of the vertices observed in the execution of the benchmark . In every level, the number of messages sent for each process is determined following Equations to .…”

“…Beamer et al characterize the memory requirements and locality of the benchmark, but they do not study the impact of the network and its utilization. Previous works from Anghel et al and Fuentes et al provide a comprehensive characterization of the communications in the simple implementation of the benchmark. Our previous work in Fuentes et al identified the strong dependence between root vertex connectivity and BFS execution and provided a first model of Graph500 traffic, which has been extended in Section 3.…”

“…The impact of the point-to-point messages to notify the end of a tree level is negligible. 7 The likewise sporadic all-reduce is important only because it synchronizes the generation of new messages. Therefore, from a network usage perspective, the major communications are the point-to-point exchanges during each tree level.…”