Leveraging Diverse Regression Approaches and Heterogeneous Machine Data in the Modeling of Computer Systems Performance

Regular forms of Amdahl’s law and the Super Serial model fail to be predictive of machine performance for heterogeneous processor datasets. In order to address this problem we successfully express Amdahl’s law and the Super Serial model in terms of common denominator processor characteristics such as threads and clock speed. The revised forms of Amdahl’s law and the Super Serial model allow leveraging heterogeneous machine data as input to capacity or scalability models. The authors choose to use regression for its formalism in verifying models. The main scientific contributions are: (1) Generalized Amdahl’s law and Super Serial model taking into account threads, clock speed, processor cores and cache size; (2) Empirical models for performance gains due to cache memory; (3) Alternatives on variable segmentation (e.g. truncating memory at saturation point). (4) A Hybrid regression method combining all the previous.