Flow control considerations in network-based architectures

In network-based parallel architectures, the issues of fairness, freedom of deadlock due to nite buuers and guaranteed message delivery can aaect not only performance but even more importantly the guarantees that the architecture provides to the users so that they can write correct, race-free portable programs. When blocking communication is assumed, the solutions to these problems can be relatively simple and eecient. When nonblocking communication must be supported for performance, these issues complicate the designs of architectures irrespective of the programming paradigm that the architecture supports and the characteristics of the network. In this paper, we consider blocking and nonblocking communication and the eeect of ow control in the latter case, on both performance and correctness. Our experimental results are derived on a cluster of dual processor SMPs connected via FDDI ber ring. The cluster is built of oo-the-shelf hardware components, although communication protocols customized for par-allelism, and in particular the remote read/write programming paradigm, have been developed and are executed by one of the processors while the other is dedicated to computation. Abstract In network-based parallel architectures, the issues of fairness, freedom of deadlock due to nite buuers and guaranteed message delivery can aaect not only performance but even more importantly the guarantees that the architecture provides to the users so that they can write correct, race-free portable programs. When blocking communication is assumed, the solutions to these problems can be relatively simple and eecient. When nonblocking communication must be supported for performance, these issues complicate the designs of architectures irrespective of the programming paradigm that the architecture supports and the characteristics of the network. In this paper, we consider blocking and nonblocking communication and the eeect of ow control in the latter case, on both performance and correctness. Our experimental results are derived on a cluster of dual processor SMPs connected via FDDI ber ring. The cluster is built of oo-the-shelf hardware components, although communication protocols customized for par-allelism, and in particular the remote read/write programming paradigm, have been developed and are executed by one of the processors while the other is dedicated to computation.