Towards asynchronous metacomputing in MPI

Metacomputing so far has been done mostly in more or less static configurations. However, applications with dynamic irregular behavior are increasing in significance and the computing platforms more often are time-sharing environments with varying system load. Thus, possibilities for dynamic connection and dynamic workload migration are becoming important. The paper discusses an approach to perform asynchronous workload balancing using the standard parallel library MPI. MPI and threads typically live in more or less separated worlds and the thread extension of MPI-2 is mainly meant to exploit more efficiently per SMP node within a model which is still mostly SPMD. We have extended MPI by dynamic mechanisms to automatically balance workload on the basis of threads and dynamic status/resource monitoring. Our extended library TeMPI is designed to run with in a minimum version with MPICH and thus MPICH-G2 in the Globus grid environment.

1 Introduction

Traditional MPI programming is focused on a basically static and mostly synchronous programming model [7]. Though, many irregular applications with a more dynamic workload can still be well supported by employing load-balancing-libraries to dynamically redistribute data in semistatic phases. And a more dynamic interaction style and asynchronous execution can still be implemented by using MPI's nonblocking communication and/or polling features. The latter means that the aplication can poll for input and dispatch between activities accordingly. This closely resembles a task-oriented style but is to be implemented at application level and awkward to write and read.