Message-passing multicomputers have been expected as the most promising way to construct massively parallel computers. The overall performance of these machines critically depends on the efficient use of system resources. Due to its partitionable structure, the multicomputers can easily support multitasking environment. In a multitasking environment, numerous tasks, each of which consists of a number of parallel modules, can be assigned to independent group of processors and executed simultaneously. Since tasks needs the various number of processors, efficient processor allocation, in order to make a multicomputer system to accommodate as many tasks as possible, is an important issue for achieving high performance on a multicomputer.
There are two processor fragmentation problem in the processor allocation, called internal and external fragmentation. These two fragmentation are occurred by topological properties of multicomputer or properties of allocation strategies. The allocation strategies are classified into two approaches, called first-fit and best-fit approaches. The best-fit approach is intended to allocate an appropriate submesh by considering the external fragmentation to reduce unnecessary fragmentation.
In this thesis, two processor allocation strategies are developed for the multicomputers. One is a best-fit submesh allocation for mesh multicomputers and the other is a fast first-fit subcube allocation for hypercube multicomputers. The proposed best-fit submesh allocation strategy is to allocate a submesh which can reserve the large free submeshes as many as possible and to maintain the high availability of the wide range of submeshes. This approach would be able to reduce the external fragmentation, especially prevent the `unnecessary' external fragmentation occurring by an inefficient allocation. To further reduction of external fragmentation, a lookahead allocation technique are proposed. We also propose the extension of the proposed strategy for faulty meshes, that is, virtual submesh allocation strategy, which can maintain and allocate virtual submeshes derived from faulty submeshes. The proposed subcube allocation strategy is the fastest recognition-complete subcube allocation strategy using cube coalescing to prune search space efficiently. The simulation results show that two strategies provide the best performance for the various performance metrics under various load and distributions.
