Bulk primitives in Linda run-time systems

This thesis investigates techniques for the efficient imple entation of the Linda parallel process coordination model for open, distributedcomputing systems. The principal focus of the research is on the use of the bulk mo ve ent of tuples within open systems which, contrary to intuition, can result in signific ant efficiency gains for a large class of problems. The emphasis onpensystems — those in which future history of process creation a nd deletion cannot be known at compile-time — is due to the curre nt interest in extending the Linda model to encompass widely distributed computing, as exempl ified by the ‘Network Computer’ notion. However, such open systems place severe constraint on the types of optimisation available relative to closed systems — in particular, the very powerfu l compile-time analysis techniques previously used are no longer feasible. Methods for the construction of efficient Linda kernels are i ntroduced based on a novel method of dynamicallyclassifying tuple spaces according to their locality, whic h allows the run-time movement of tuple spaces’ locations within the distributed kern el. An important consequence of the proposed technique is that it does not require any ‘global’ i nformation — it works solely on information locally available to each component of the distributed kernel. Equa lly importantly, the scheme is entirely transparent to the programmer, and there for requires no user-supplied ‘hints’ or ‘pragmas’. The implemented kernel is fully distributed, consequently tuples within a particular tuple space may be stored on several physical nodes. The kernel supports standard Linda with multiple tuple spaces, thecollect primitive, and another primitive called copy-collect. The justification for the addition ofcopy-collect is the multiplerd problem which is described in detail in this dissertation. No acceptable way of overcoming the mult iple rd problem without the use of thecopy-collect primitive has been published. The performance of the implemented kernel is shown to be sign ifica tly better than the performance of a kernel that does not use the bulk movement of tuples , and through using a “real-world” example the kernel is shown to provide, under some circumsta nces, better performance than the best commercially available closed implementation which u ses compile time analysis. Finally, an extension of the concept of classifying tuple sp aces is presented, which generalises the concept leading to a detailed proposal for a multi-layer hi archical kernel, which is more scalable than current traditional implementations.