Successful deployment of global computing grids involves solving a number of critical issues, including authorization, security, and, last but not least, the communication latency and limited bandwidth. The performance of computational grids is especially affected by network communication delays. Over the years, the developers of grid middleware and applications have come up with different approaches to hide the latency of the network to improve the performance and scalability.

The three main approaches to improve the performance of the trivial synchronous model are bulk communication, the asynchronous model, and pipelining. This paper describes the models in depth, including the equations for determining the communication time for each model. The authors do not limit themselves to the theory, but also present the results of their experiments measuring the performance of the execution models in local and wide area network scenarios.

Speaking of the test results, I was surprised to see the poor performance of the pipeline model in the wide area network throughput benchmark. The authors’ explanation is that this is due to the network transport time being the largest component of the execution time in this benchmark. This is an acceptable explanation, but I would still like to see this benchmark executed for a larger number of operations and smaller response sizes. My other complaint is that the figure showing the performance of the pipeline as the function of the pipeline depth does not seem to match the text describing the data in the chart.

As for the conclusions, the first one is rather obvious: which model performs best depends on the given scenario, the type of application, and the network performance. Much more interesting is the second conclusion that all of the models can be combined into one generalized model. Drawing from this conclusion, the authors mention their reference implementation of the SAGA application programming interface—a high-level middleware designed to allow the execution model to be changed simply by changing some parameter settings.

Despite the objections above, I found this paper well written, educational, and certainly worth the time I spent poring over the benchmark results.