Predictive Evaluation of Partitioning Algorithms through Runtime Modelling

Performance modelling unstructured mesh codesis a challenging process, due to the difficulty of capturing theirmemory access patterns, and their communication patterns atvarying scale. In this paper we first develop extensions to anexisting runtime performance model, aimed at overcoming theformer, which we validate on up to 1,024 cores of a Haswell-based cluster, using both a geometric partitioning algorithmand ParMETIS to partition the input deck, with a maximumabsolute runtime error of 12.63% and 11.55% respectively. Toovercome the latter, we develop an application representative ofthe mesh partitioning process internal to an unstructured meshcode. This application is able to generate partitioning data thatis usable with the performance model to produce predictedapplication runtimes within 7.31% of those produced usingempirically collected data. We then demonstrate the use of theperformance model by undertaking a predictive comparisonamong several partitioning algorithms on up to 30,000 cores. Additionally, we correctly predict the ineffectiveness of thegeometric partitioning algorithm at 512 and 1024 cores