S. Juh´asz
Asynchronous communication, cluster of workstations, parallel com-munication channels, performance modelling
The overall performance characteristics of cluster systems depend heavily on the pattern and on the amount of communication between the nodes. The performance may be improved by using asynchronous (nonblocking) message passing, because it allows communication and computation to overlap, thereby hiding a part of the communication overhead. This paper develops an analytical model to capture the performance-related issues of asynchronous communication in a small, fully switched cluster environment. The parameters of the model can be identified from measurable program and hardware characteristics, allowing the model to anticipate the performance behaviour of complex parallel applications. The paper’s main contribution is to describe the effect of parallel communication channels on the effective bandwidth of a single node. The model is validated by comparing the predicted and measured performance of two different broadcast primitives for a range of message sizes as a function of the number of the participating nodes.
Important Links:
Go Back
