Think globally, search locally

DMCA

Think globally, search locally

Cached

Download Links

Other Repositories/Bibliography

DBLP

by Kamen Yotov , Keshav Pingali , Paul Stodghill

Venue:	In ICS ’05: Proceedings of the 19th Annual International Conference on Supercomputing
Citations:	22 - 0 self

BibTeX

@INPROCEEDINGS{Yotov_thinkglobally,,
    author = {Kamen Yotov and Keshav Pingali and Paul Stodghill},
    title = {Think globally, search locally},
    booktitle = {In ICS ’05: Proceedings of the 19th Annual International Conference on Supercomputing},
    year = {},
    pages = {141--150}
}

OpenURL

Abstract

A key step in program optimization is the determination of optimal values for code optimization parameters such as cache tile sizes and loop unrolling factors. One approach, which is implemented in most compilers, is to use analytical models to determine these values. The other approach, used in library generators like ATLAS, is to perform a global search over the space of parameter values by generating different versions of the code and executing them on the actual machine to find the parameter values that give the best performance. Neither approach is suitable for use in general-purpose compilers that must generate high quality code for large programs running on complex architectures. Model-driven optimization may incur a performance penalty of 10-20 % even for a relatively simple code like matrix multiplication, as was shown recently by Yotov et al. On the other hand, global search is not tractable for optimizing large programs for complex architectures because the optimization space is too large. To address this problem, some researchers are exploring more sophisticated search algorithms such as the simplex method, but it remains to be seen if these methods are successful in reducing search time without compromising on the quality of the solution. In this paper, we advocate a different methodology for generating high-performance code without increasing search time dramatically. Our methodology has three components: (i) modeling, (ii) local search, and (iii) model refinement. We use analytical models to estimate optimal values for transformation parameters. Since it is impossible to build tractable analytical models that capture all the features of

Keyphrases

global search optimal value complex architecture search time large program analytical model parameter value optimization space different methodology sophisticated search algorithm loop unrolling factor key step tractable analytical model simple code high quality code library generator code optimization parameter matrix multiplication transformation parameter model refinement local search cache tile size simplex method program optimization general-purpose compiler high-performance code actual machine different version performance penalty model-driven optimization

DMCA