Efficient Management of Huge Data Sets on Cluster Computers

In a cluster computer a parallel file system is encharged to spread one single parallel file on the different computer's I/O nodes using a determined distribution function. In file I/O intensive parallel scientific applications with "semi-random temporal parallel file I/O acess patterns", this file is accessed at different addresses at the sametime by a number of processes that may vary between two consecutive iterations. In this thesis a set of "semi-random temporal parallel file I/O access patterns" generated by a phylogenetical application is categorized. For these patterns a partitioning function is proposed that guarantees at any time during execution access to the parallel file. This thesis shows the correlation existing between the type of I/O access patterns and the type and setting of two round robin based distribution functions so that the overall application's execution time can be reduced