Open AccessMoving towards grey‐box predictive models at micro‐architecture level by investigating inherent program characteristics
Author(s) -
Ahmadinejad Hoda,
Fatemi Omid
Publication year - 2018
Publication title -
iet computers and digital techniques
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.219
H-Index - 46
ISSN - 1751-861X
DOI - 10.1049/iet-cdt.2016.0148
Subject(s) - computer science , process (computing) , similarity (geometry) , set (abstract data type) , scheme (mathematics) , space (punctuation) , machine learning , predictive modelling , data mining , artificial intelligence , mathematics , mathematical analysis , image (mathematics) , programming language , operating system
Predictive modelling has gained much attention in the last decade, aiming fast evaluation of different design points in design space exploration (DSE) process. However, predictive model construction still requires costly simulations for every unseen program. To reduce the number of simulations, several cross‐program prediction schemes have been developed. This study proposes a cross‐program predictive scheme for micro‐architectural DSE. The scheme measures a set of representative inherent characteristics for the unseen program and compares them against the same characteristics of training programs. Then, based on similarity information, the performance trend of the unseen program is predicted using the predictive models of training programs. As the raw data of the characteristics do not characterise programs in performance space, the authors propose a novel method which transforms the characteristic space into performance space. The proposed method achieves 13.3× speed‐up over the program‐centric scheme with an average correlation coefficient of 0.92.