Automated Design of Computer Clusters

Computer clusters made of mass-produced, off-the-shelf components are able to deliver the same real-life performance on a variety of tasks as other types of supercomputers, but at a lower cost. Current ad hoc design practices are characterised by only partial exploration of design space and inability to accurately predict capital and operating expenses. We provide the framework for a more comprehensive design space exploration, paving the way to a future CAD system for computer clusters and warehouse-scale computers with decision support and “what if” analysis capabilities. We formulate the design task as a discrete combinatorial optimisation problem, with the non-linear objective function being the ratio of total cost of ownership to real-life performance. Although complex, the objective function allows for an unbiased assessment of proposed design alternatives. Various constraints can be imposed on technical and economic characteristics of the computer cluster, including minimal performance, maximal capital and operating expenses, power consumption, occupied space, etc. Obtaining the value of the objective function is only possible after several consecutive stages of design process are completed: choosing an optimal configuration of a single compute node, designing interconnection network and a power supply system, etc. We provide a prototype CAD tool that implements these stages. We also propose heuristics to deal with combinatorial explosion at various stages.