Parameter Space Exploration：Case Study on Compiler Options
Date Issued
2009
Date
2009
Author(s)
Chen, Jian-Hua
Abstract
Optimizing computer systems is essential. However, modern computer systems are complex. A key driver of performance is parameter setting. There are many tunable parameters and optimizing them is not easy. With careful selection of the parameter values, better performance and higher utilization can be achieved. In this thesis, we consider a well-recognized tunable parameter set, the compilation parameters.ompilers provide many optimization options as part of their APIs. It is hard even for a compiler expert to predict which optimization options should be turned on for optimizing a specific application. In addition to turning on or off an optimization, it might be useful if we consider the setting of some parameters that affect the behavior of the optimizations. Therefore, we proposed a multi-objective compiler option searching approach based on genetic algorithm to fast locate options that improve the applications. We achieve 11.8% speedup on SPEC CPU2006 programs over the most aggressive optimization configuration of the GNU Compiler Collection (GCC) compiler (-O3) by tuning the parameter setting of the feedback-directed optimization options of GCC. Also, we reduce the code size of the native codes on Android system by 6.6% against its default configuration.
Subjects
Compiler
Optimization
Genetic Algorithm
Parameter
Type
thesis
File(s)![Thumbnail Image]()
Loading...
Name
ntu-98-R96922095-1.pdf
Size
23.32 KB
Format
Adobe PDF
Checksum
(MD5):1048e562b0d519fdaad3d0472950fcc9