Files
Abstract
Scalability prediction is one of the key problems facing high performance computing today. Methods to predict scalability accurately are necessary in order to improve throughput and overall efficiency on large-scale machines. This dissertation presents our novel, regression-based system for accurately predicting the scalability of scientific applications on large-scale machines. Our regression-based system provides accurate runtime predictions on large processor counts for multiple scientific applications when run using strong scaling. Our system is also able to provide input parameters leading to accurate time-constrained scaling on larger processor counts. We also discuss the impact of noise on scalability prediction. This work takes large steps towards a general scalability prediction system that could be deployed on supercomputing systems in the near future.