The Computer as a Solution Space

Abstract

Zenotech provides cloud-based high performance computing (HPC) ser- vices to customers wanting to run large parallelised simulations. The transfer of data between nodes in a parallelised problem takes a few orders of magnitude more time than the calculation of the data itself, causing a bottleneck. Any improvements to algorithm implementation that can improve efficiency, reliability, accuracy or robustness to errors would mean more effective and economic use of HPC resource. Zenotech challenged the study group to find a mathematical characterisation of a HPC system that would allow for a more effective mapping of algo- rithm topology onto HPC network topology and a better understanding of scalability and robustness in the system. We investigated how to best optimise partition size and time of data transfer on a square lat- tice and the effect of using old boundary data for continued calculation when new data does not arrive on time. We also produced a toy network model to test the effect of different partition choices and communication strategies.