DMCA
Page 1 of 10 Comparing and Combining Profiles (1999)
| Venue: | In Second Workshop on Feedback-Directed Optimization (FDO |
BibTeX
@INPROCEEDINGS{And99page1,
author = {Serap Savari And and Serap Savari and Cliff Young},
title = {Page 1 of 10 Comparing and Combining Profiles},
booktitle = {In Second Workshop on Feedback-Directed Optimization (FDO},
year = {1999},
pages = {50--62}
}
Share
 |
 |
 |
 |
||
OpenURL
Abstract
this paper to branch profiles; much of the relevant scholarly work also relates to branch profiling. Our first two problems, comparing profiles and determining when program behavior has changed, suggest that we need some sort of metric of the differences between profiles. Prior work in this area includes the concepts of coverage, conflict, and a geometric measure introduced recently by Kistler and Franz [8]. In their extensive cross-validation study of static branch prediction, Fisher and Freudenberger discuss coverage, noting that some runs of the program execute different parts of the program from other runs [6]. They experimented with a number of different metrics but did not find any one that satisfactorily explained their results. Calder, Grunwald, and Srivastava later defined coverage as the percentage of branches from a program run that were also executed during a profile training run [3]. We adopt the latter definition of coverage for this paper, and we define both static coverage and dynamic coverage. Static coverage is the number of branch instructions touched in both the training and testing runs divided by the number of branch instructions touched during the testing run. Dynamic coverage is similar, but it weighs each branch by its execution frequency. The dynamic coverage is the sum of execution frequencies of instructions in the testing run that were also touched during the training run divided by the total branch execution frequency of the testing run
Keyphrases
dynamic coverage combining profile branch instruction testing run execution frequency static coverage prior work training run static branch prediction geometric measure program behavior program execute different part program run total branch execution frequency relevant scholarly work different metric extensive cross-validation study latter definition profile training freudenberger discus coverage
