Abstract-This paper presents the performance analysis of a grid-enabled MIP solver in a grid environment consisting of three clusters on a campus LAN. In particular, the paper focuses on the analysis of the behavior of the application using two networks connecting the clusters with different latency and bandwidth.
I. INTRODUCTIONThe goal of computational grids is to aggregate numerous heterogeneous resources in order to provide sufficient computational power to solve large-scale problems. In a grid environment, users can use unlimited resources as a single powerful entity. However, the heterogeneity of the resources in the grids makes it difficult to predict the performance behavior of the applications. The performance analysis of an application on a grid is a very interesting matter, because the results are dependent on the problem, on the code structure and the communication patterns exploited, on the performance of the grid hardware used (computing node, CPU, memory hierarchy), on the middleware and the run-time systems used, and, above all, on the communication performance inside and among the computational facilities making up the grid. This paper tackles the grid performance analysis problem for a real-world code, a solver of Mixed Integer Programming problems (MIP) developed using the BCP-G framework. BCP-G is a customized version of COIN/BCP, an open source framework developed within the IBM COIN-OR project [1], which is based on the Branch, Cut and Price method for solving large-scale linear integer programming problems. The original COIN/BCP framework, based on the use of PVM libraries [2], has been provided with a new MPI communication API able to exploit the MPICH-G2 system, a grid-enabled MPI implementation [3], [4]. Moreover, the developed application is part of a more complex grid-enabled system, consisting of another framework (Meta-Pbc) and a web portal (SWI-Portal). The first one is a new framework, based on a decentralized master/worker schema [5]. The second one is a web portal designed to manage users and jobs as user-friendly as possible.In a previous paper [6] we studied the performance of the software system sketched above, on a test grid environment consisting of three small clusters on a fairly slow campus LAN. Even if the use of the three clusters led to a nonnegligible speedup, the performance gain was relatively small compared to the one that could be obtained using a single