COBAYN is partially funded by ANTAREX European project.
All rights reserved for the author.
Any non-academic usage of COBAYN must be permitted by the author.
v1.0 July 2017
Amir H. Ashouri
[email protected]
This is a minimal working version of the COBAYN approach. It is a Matlab project that predicts the best set of compiler flags given a new unseen application. Evaluations showed that it can outperform the existing models, i.e., [Agakov et al. 2006], [Park et al. 2015], and GCC's standard optimization levels -O2 and -O3. COBAYN tackles the problem of selecting the best compiler passes, it does not consider the phase-ordering problem. See my other publications for papers related to finding the best ordering of optimization phases.
For the details of the methodology, COBAYN can be found at ACM digital library.
If you use any of the materials in this project, i.e., codes, provided ready-to-go datasets, methodology, etc., you should cite this work:
*Bibtex version:
@article{Ashouri2016COBAYN,
author = {Ashouri, Amir Hossein and Mariani, Giovanni and Palermo, Gianluca and Park, Eunjung and Cavazos, John and Silvano, Cristina},
title = {COBAYN: Compiler Autotuning Framework Using Bayesian Networks},
journal = {ACM Trans. Archit. Code Optim.},
issue_date = {June 2016},
volume = {13},
number = {2},
month = jun,
year = {2016},
issn = {1544-3566},
pages = {21:1--21:25},
articleno = {21},
numpages = {25},
url = {http://doi.acm.org/10.1145/2928270},
doi = {10.1145/2928270},
acmid = {2928270},
publisher = {ACM},
address = {New York, NY, USA},
keywords = {compiler optimization, Bayesian networks, machine learning, statistical inference},
}
*ACM version:
Amir Hossein Ashouri, Giovanni Mariani, Gianluca Palermo, Eunjung Park, John Cavazos, and Cristina Silvano.
2016. COBAYN: Compiler Autotuning Framework Using Bayesian Networks.
ACM Trans. Archit. Code Optim. 13, 2, Article 21 (June 2016), 25 pages.
DOI: http://dx.doi.org/10.1145/2928270
|── COBAYN/
| ├── COBAYN.m // automated script ro run and evaluate COBAYN
| ├── README.txt
| ├── data
| │ ├── cBench_onPandaboard_24app_5ds.csv // Execution-times dataset
| │ ├── ft_MICA_cbench.csv // static characterization dataset
| │ └── ft_Milepost_cbench.csv // Dynamic characterization dataset
| ├── dataProcessing
| │ ├── getBestSet.m
| │ ├── getBestSet2.m
| │ └── take_db.m
| ├── initMatEnv.m
| ├── model
| │ ├── enterEvidence.m
| │ ├── generateModels.m
| │ ├── getMLE.m
| │ ├── getModel.m
| │ ├── sampleModel.m
| │ └── useModels.m
| |── utils
| | ├── cell2csv.m
| | ├── pcaFromStatToolbox.m
| | |── predictionTableGenerator.m
├── Estimation_of_Distribution_Algorithms (*) // Original Matlab BN package
│ ├── BNT
│ ├── BNT_SLP
│ ├── Mateda2.0
│ └── setupMatlabEnv_EDA.m
└── README.md
In order to facilitate the test and future comparisons, I have provided a dataset consisting of execution times for many application and two different characterization methods: (i) an exploration dataset, and, (ii) two different characterization datasets provided ready-to-go. They can be found at the ~COBAYN/COBAYN/data directory.
*(i) Exploration dataset: It consists of 7 compiler passes (GCC flags) (mentioned in the COBAYN's paper) with 24 applications taken from Cbench suite. Each application has been evaluated with 5 different datasets for its execution time and code-size. They are profiled in a Pandaboard embedded device running on Arch-linux.
*(ii) Characterization datasets are derived with MICA (dynamic instrumentation) and Milepost (static characterization).
-Please set your absolute paths in these files:
COBAYN/Estimation_of_Distribution_Algorithms/Mateda2.0/InitEnvironments.m
COBAYN/Estimation_of_Distribution_Algorithms/BNT_SLP/add_SLP.m
-Change this line to your needs:
COBAYN/COBAYN/model/generateModels.m:25
bestPerBenchmark = X; // use X number of the best samples (sorted by execution time) in each application
A quick automated script that takes the DBs (both the exploration and the characterization), does the import, train, and prediction (cross-validation) is available by just running the following command:
>>cd COBAYN
>>COBAYN.m
Modify the "take_db.m" file located at ~/BeliefNetwork/dataProcessing/take_db.m based on you need as you have to mention the following:
1- Importing files:
- Exploration Data set (default is cBench_onPandaboard_24app_5ds.csv)
- Application characterization file (default is ft_Milepost_cbench.csv)
2- Dimension Reduction Technique
- Default is PCA
3- Dataset Normalization Technique
- pick 1-5 and comment the others
*Exploration Data set: Your input dataset.csv must be a N columns csv file and conformed to the following structure:
FIRST_ROW = Header
COL#{1} = APP_NAME
COL#{2:M} = COMPILER_FLAGS (binary (categorical) flags: 0 _or_ 1)
COL#{M+1:N-1} = EXEC_TIME per each datasets (Numbers)
COL#{N} = CODE_SIZE
- Application characterization
You application characterization.csv set must be n columns csv and have the following structure:
FIRST_ROW = Header
COL#{1} = APP_NAME (i.e. automatic_bitcount)
COL#{2} = DATA_SET_No (i.e: dataset1)
COL#{3:n} = FEATURES (Number)
In order to train the COBAYN BN network, you should follow these steps:
>>init_MatEnv.m // setup the paths
>>take_db.m // imports the dataset
>>generateModels.m // trains COBAYN
Cross-validation Mode
>>predictionTableGenerator.m // predicts per application and generates --> Generates COBAYN prediction matlab structure
Standalone Test
Currently not available.
The Bayesian network package is dervided from the work of "MATEDA-2.0":
@article{santana2010mateda,
title={Mateda-2.0: Estimation of distribution algorithms in MATLAB},
author={Santana, Roberto and Bielza, Concha and Larranaga, Pedro and Lozano, Jose A and Echegoyen, Carlos and Mendiburu, Alexander and Armananzas, Rub{\'e}n and Shakya, Siddartha},
journal={Journal of Statistical Software},
volume={35},
number={7},
pages={1--30},
year={2010}
}
All rights reserved for the author.
Any non-academic usage of COBAYN must be permitted by the author.
v1.0 July 2017
Amir H. Ashouri
[email protected]