A message passing interface library for inhomogeneous coupled clusters

Abstract

Clusters of PC containing mostly general purpose hardware have become more and more usable for high performance computing tasks in the past few years. Clustering existing systems or extending cluster installations leads to the problem of inhomogeneous cluster installations with a broad variety of hardware and even operating systems. This paper describes the efforts of the MetaMPICH project to deal with multi-platform inhomogeneous clusters, using different networks and processor types. An approach is presented which on one hand hides the heterogeneity of softand hardware from the MPI application programmer, but on the other hand supports the exploitation of special abilities of single meta computing resources by providing a new MPI communicator which reflects the real topology of the connected systems.

Hierarchy-Aware Message-Passing in the Upcoming Many-Core Era

“…For that reason their application area is also referred to as a so-called meta-computing environment. The most common Grid-enabled MPI libraries are MPICH-G2 (Karonis et al, 2003), PACX-MPI (Gabriel et al, 1998), GridMPI (Matsuda et al, 2004), StaMPI (Imamura et al, 2000), MPICH/Madeleine (Aumage et al, 2001) and MetaMPICH (Pöppe et al, 2003), which are all proven to run large-scale applications in distributed environments. Although these meta-MPI implementations usually use native MPI support for site-internal communication, as for example provided by a site-local vendor MPI, they must also be based on at least a transport layer being capable of wide area communication for bridging and forwarding messages also to the remote sites.…”

Design and Implementation of a Service-integrated Session Layer for Efficient Message Passing in Grid Computing Environments

“…However, as a first feasibility study, we have already successfully integrated ISI's functionalities into a project called MetaMPICH, which is about the implementation and enhancement of a Grid-enabled MPI library and which has also been conducted at our institute. [21] In order to meet the demands of heterogeneous Grid environments, this MPI library provides even two different methods for linking the remote sites: a router-based method and a so-called multi-device method [3]. In particular the multi-device method of MetaMPICH has significantly benefited from an ISI managed session establishment.…”

A Fair Benchmark for Evaluating the Latent Potential of Heterogeneous Coupled Clusters

“…Since MPI [20] is the most important API for implementing parallel programs, several MPI libraries have been extended to meet the demand of distributed and heterogeneous computing. The most common meta-computing and grid-enabled MPI libraries are MPICH-G2 [17], PACX-MPI [5], MPICH/Madeleine [2], GridMPI [18], StaMPI [25] and MetaMPICH [23], which are all proven to run large-scale applications on coupled cluster systems.…”