Abstract | An attempt is made to understand the properties of I/O between compute nodes in a parallel system. A theoretically plausible framework to emulate an application's I/O communication pattern (without implementation) is presented. The focus, however, lies in a literature review, in conjunction with a set of performance experiments (using LAM, a known MPI implementation, and mpptest), which quantify and expose some of the characteristics of the communication within two radically different Beowulf clusters. |
---|