The 7th Annual Meeting on High Performance Computing and Infrastructure in Norway

NOTUR2008 - ABSTRACTS

Comparing the Performance of Multiple Single-Cores vs. a Single Multi-Core

Otto Anshus

Moore's law states that the number of transistors per inch2 doubles every 18 months. The doubling of transistors has been used to double the performance, and to increase the functionality of the processors. However, power consumption and heat issues, limited instruction-level parallelism and increased complexities of the processors have stimulated a rapid shift to a situation where the doubling of transistors are now being used to double the number of cores. Modern graphics processors (GPUs) are multi-core, and they are emerging as an interesting platform for parallel, compute-intensive general-purpose computation. However, it is not simple to deliver the performance to the application level software. One challenge is to map the threads and processes of the application to the cores. Another is the lack of strong synchronization primitives. This talk compares the performance of the embarrassingly parallel Mandelbrot set computation on four systems: a single single-core computer, a cluster of 28 single-core computers, a cluster of 28 single-core computers where the main processing is done on older multi-core graphics card on each computer, and a single single-core computer with a modern multi-core graphics card. The results show that a single muli-core graphics card can significantly outperform a cluster of 28 computers and slightly older graphics cards. This is explained by the high data-parallelism and extreme thread-level parallelism provided by the GPU.

< Back to programme