High performance communications for hypercube multiprocessors.

A large potential exists for increasing the communication performance of hypercube multiprocessors. The development of this potential will allow an increasing variety of algorithms to be efficiently executed on these systems. In this dissertation we introduce a classification scheme for parallel programs and their communications. The classification scheme establishes a framework to facilitate discussion of communication system issues. This framework is then used to identify classes of programs with similar communication characteristics. The exploration of communication issues is divided into two broad groups: communications among neighboring nodes or broadcasts to all other nodes, and communications among non-neighboring nodes. We show that the execution time of many programs that employ communications from the first group may be improved considerably by utilizing a new operating system communication module that we have developed and implemented. We then describe the minor architectural support that is necessary to allow the remainder of the programs that use broadcast or neighboring communications to take advantage of our improved communication system. Finally, we develop a new routing scheme that reduces the communication latency for communications among non-neighboring nodes. As a proof of principle, we provide a detailed description of a communication architecture that employs this routing scheme and also allows greater overlap of computations with communications. Our simulations of this new architecture predict significant performance improvements for non-neighboring communications as well as further improvements to neighboring and broadcast communications.Ph.D.Computer scienceElectrical engineeringUniversity of Michiganhttp://deepblue.lib.umich.edu/bitstream/2027.42/161904/1/8821553.pd