Comparing the performance of high-level middleware systems in shared and distributed memory parallel environments

The utilization of toolkits for writing parallel and/or distributed applications has been shown to greatly enhance developer's productivity. Such an approach hides many of the complexities associated with writing these applications, rather than relying solely on programming language aids and parallel library support, such as MPI or PVM. In this work, we evaluate three different middleware systems that have been used to implement a computation and I/O-intensive data analysis application from the domain of computer vision. This study shows the benefits and overheads associated with each of the middleware systems, in different homogeneous computational environments and with different workloads. Our results lead the way toward being able to make better decisions for tuning the application environment, for selecting the appropriate middleware, and also for designing more powerful middleware systems to efficiently build and run highly complex applications in both parallel and distributed computing environments.