I/O Processor Allocation for Mesh Cluster Computers

As cluster systems become increasingly popular, more and more paralle applications require need not only computing power but also significant I/O performance. However, the I/O subsystem has become the bottleneck of the overall system performance for years due to slower improvement of the second storage devices. In recent years parallel I/O has drawn an increasing attention as a promising approach
to eliminate this bottleneck. To improve I/O efficiency of a cluster system computation tasks must be carefully assigned to processors, so that the communication overheads within the group the processors of
the task, and those I/O traffics that connect processors of the task to I/O system are both optimized. Earlier processor allocation strategies considered the optimization of communication traffic or I/O
traffic only. Since both the communication and I/O traffic can cause network contention, we develop two groups of algorithms -- binary tree based methods and Snake-Hilbert curve based methods, that address the issues of both communication and I/O traffics simultaneously. The experimental results indicate that for tasks that
have different mixture of communication and I/O traffics, our algorithms have very good performance in terms of overall parallel I/O efficiency. We also developed two mathematical evaluating criteria -- ``compactness' and ``spatial compactness', to determine the fitness of allocation algorithms in terms of geometrical adjacency of
processors. The theoretical results of these two criteria are also presented in this dissertation.