Towards modeling and integrated design automation of supercomputing clusters (MIDAS)

Abstract

The choice of supercomputers, by the user community, should not merely be based on the benchmarks. With the advent of supercomputers delivering petaops performance, computationally intensive applications such as brain modeling and energy requirement prediction have become predominant. Conventional benchmarks do not reflect the functional characteristics of these applications. Thus, the gap between the performance projected and the actual performance delivered when the application is ported onto the cluster is wide. The primary cause for this disparity in performance is due to the user community’s lack of knowledge about the system and improper design choices. With the supercomputer market being seller biased this problem is worsened. Design and development of very large systems like the clusters involve massive efforts and financial investment. Creating a design automation environment for the clusters is bound to bring down the man years and hence the production cost. Currently there are adhoc approaches which lack an integrated design space to cater to the wider needs of the user community. Thus, an integrated design framework that takes all the components of the supercomputing cluster and the relationship across the design spaces is essential. In this paper we propose a novel methodology called the Modeling and Integrated Design Automation of Supercomputing Cluster (MIDAS). MIDAS is the birth of a new design philosophy which requires lot of research and tuning, providing scope for evolution and optimization.