Section: New Results

Affinity Abstraction

This year we worked on the affinity abstraction. Often, the affinity between two processes or threads is measured by the a matrix where a high entry represent a high affinity. Such example of matrices are the number of messages and the number of bytes exchanged between processing units. However, such matrix hide many characteristics of the application such as computation/communication overlap, network contention, etc. First, we have developed a new OpenMPI PML module to gather communication matrix of a running application. Then, we have conducted an extensive study of the minighost application to understand how such communication matrix actually measure the affinity between processes. On this application it appears that the size metric better matches the performances and that the performance of process placement is highly correlated to the proportion of communication in the application.