Interconnection networks performance modelling for multi-core multi-cluster architecture

Abstract

In High Performance Computing (HPC) system design and deployment there is an increasing trend towards networked parallel systems such as cluster computing systems. Cluster computing is typically built from a group of workstations connected by high-speed networks to form a single high-availability system. One of the driving forces behind high-performance clusters is the advent of multi-core clusters. The aim of the research reported here is to design a new architecture for large-scale multi-core cluster computing systems and to investigate the interconnection network performance of the new architecture.

Since the overall performance of cluster computing systems always depends on the efficiency of its communication networks, performance analysis of the interconnection networks is vital. A general problem in the network may arise from the fact that multiple messages can be in transmission at the same time, using the same network links.

The contribution of this thesis is to develop a new architecture known as Multi-core Multi-cluster Architecture (MCMCA), composed of numbers of clusters where each cluster is a multi-core processor. Next, a simulation model is built to investigate the interconnection network performance of the new architecture, and the results are presented. The main performance metrics to be simulated are the latency and network throughput. The model is then used to evaluate the impact on scalability and cluster size of the interconnection network performance. Finally, analytical model including statistical analysis are used to validate the simulation results under various working conditions.

The analysis indicates that, from single-core to multi-core, there is a significant improvement in processor performance. To judge from the latency results, compared to single-core cluster a multi-core cluster can improve the network performance. Another observation is that the architecture can achieve lower latency and higher throughput as the number of cores increases. The experiments also demonstrated that a multi-core cluster can scale better than a single-core cluster. The results comparison between the analytical model and those produced from the simulation experiments has shown that the derived simulation model provides a good basis for predicting the communication delay of the interconnection network performance of the Multi-Core Multi-Cluster Architecture (MCMCA).

More information

Identifiers

Catalogue record

Export record