BinTuner is a cost-efficient auto-tuning framework, which can deliver a near-optimal binary code that reveals much more differences than -Ox settings.

BinTuner

BinTuner is a cost-efficient auto-tuning framework, which can deliver a near-optimal binary code that reveals much more differences than -Ox settings. it also can assist the binary code analysis research in generating more diversified datasets for training and testing. The BinTuner framework is based on OpenTuner, thanks to all contributors for their contributions.

The architecture of BinTuner:

The core on the server-side is a metaheuristic search engine (e.g., the genetic algorithm), which directs iterative compilation towards maximizing the effect of binary code differences.

The client-side runs different compilers (GCC, LLVM ...) and the calculation of the fitness function.

Both sides communicate valid optimization options, fitness function scores, and compiled binaries to each other, and these data are stored in a database for future exploration. When BinTuner reaches a termination condition, we select the iterations showing the highest fitness function score and output the corresponding binary code as the final outcomes.

System dependencies

A list of system dependencies can be found in packages-deps which are primarily python 2.6+ (not 3.x) and sqlite3.

On Ubuntu/Debian there can be installed with:

sudo apt-get update
sudo apt-get upgrade
sudo apt-get install `cat packages-deps | tr '\n' ' '`

Installation

Running it out of a git checkout, a list of python dependencies can be found in requirements.txt these can be installed system-wide with pip.

sudo apt-get install python-pip
sudo pip install -r requirements.txt

If you encounter an error message like this:

Could not find a version that satisfies the requirement fn>=0.2.12 (from -r requirements.txt (line 2)) (from versions:)
No matching distribution found for fn>=0.2.12 (from -r requirements.tet (line 2))

Please try again or install each manually

pip install fn>=0.2.12
...
pip install numpy>=1.8.0
...

If you encounter an error message like this:

ImportError: No module named lzma

Please install lzma

sudo apt-get install python-lzma

If you encounter an error message like this:

assert compile_result['returncode'] == 0
AssertionError

Please confirm how to use the compiler in your terminal, such as GCC or gcc-10.2.0 it needs to be modified in your .Py file

If you encounter an error message like this:

sqlalchemy.exc.OperationalError: (pysqlite2.dbapi2.OperationalError) database is locked [SQL: u'INSERT INTO tuning_run (uuid, program_version_id, machine_class_id, input_class_id, name, args, objective, state, start_date, end_date, final_config_id) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)'] [parameters: ('b3311f3609ff4ce9aa40c0f9bb291d26', 1, None, None, 'unnamed', 
   
   
    
    , 
    
    
     
     , 'QUEUED', '2021-xx-xx 03:42:04.145932', None, None)] (Background on this error at: http://sqlalche.me/e/e3q8)

Just delete the DB file saved before (PATH:/examples/gccflags/opentuner.db/Your PC's Name.db).

Install Compiler

GCC

Check to see if the compiler is installed

e.g.

gcc -v  shows that
gcc version 7.5.0 (Ubuntu 7.5.0-3ubuntu1~18.04)

Please note that there have different optimization options in different versions of compilers.

If you use the optimization options that are not included in this version of the compiler, the program can not run and report an error.

It is strongly recommended to confirm that the optimization options are in the official instructions of GCC or LLVM before using them.

e.g. GCC version 10.2.0.

You can also use the command to display all options in terminal

gcc --help=optimizers


The following options control optimizations:
  -O
   
   
    
                      Set optimization level to 
    
    
     
     .
  -Ofast                      Optimize for speed disregarding exact standards
                              compliance.
  -Og                         Optimize for debugging experience rather than
                              speed or size.
  -Os                         Optimize for space rather than speed.
  -faggressive-loop-optimizations Aggressively optimize loops using language
                              constraints.
  -falign-functions           Align the start of functions.
  -falign-jumps               Align labels which are only reached by jumping.
  -falign-labels              Align all labels.
  -falign-loops               Align the start of loops.
  ...

LLVM

clang -v

Check how to install LLVM here

https://apt.llvm.org/

https://clang.llvm.org/get_started.html

Checking Installation

Enter the following command in terminal to test:

[email protected]:~/BinTuner/examples/gccflags$ python main.py 2

You will see some info like this:

Program Start
************************ Z3 ************************
5- Result--> Unavailable
3- Result--> Available
[ Z3 return Results = first second True four False]
[ Changed "shrink-wrap" value ]
...
-------------------------------------------------

--- BinTuner ---
--- Command lines and compiler optimization options ---:
gcc benchmarks/bzip2.c -lm -o ./tmp0.bin -O3 -fauto-inc-dec -fbranch-count-reg -fno-combine-stack-adjustments 
-fcompare-elim -fcprop-registers -fno-dce -fdefer-pop -fdelayed-branch -fno-dse -fforward-propagate -fguess-branch-probability 
-fno-if-conversion2 -fno-if-conversion -finline-functions-called-once -fipa-pure-const -fno-ipa-profile -fipa-reference 
-fno-merge-constants -fmove-loop-invariants -freorder-blocks -fshrink-wrap -fsplit-wide-types -fno-tree-bit-ccp -fno-tree-ccp 
-ftree-ch -fno-tree-coalesce-vars -ftree-copy-prop -ftree-dce -fno-tree-dse -ftree-forwprop -fno-tree-fre -ftree-sink -fno-tree-slsr 
-fno-tree-sra -ftree-pta -ftree-ter -fno-unit-at-a-time -fno-omit-frame-pointer -ftree-phiprop -fno-tree-dominator-opts -fno-ssa-backprop 
-fno-ssa-phiopt -fshrink-wrap-separate -fthread-jumps -falign-functions -fno-align-labels -fno-align-labels -falign-loops -fno-caller-saves 
-fno-crossjumping -fcse-follow-jumps -fno-cse-skip-blocks -fno-delete-null-pointer-checks -fno-devirtualize -fdevirtualize-speculatively 
-fexpensive-optimizations -fno-gcse -fno-gcse-lm -fno-hoist-adjacent-loads -finline-small-functions -fno-indirect-inlining -fipa-cp 
-fipa-sra -fipa-icf -fno-isolate-erroneous-paths-dereference -fno-lra-remat -foptimize-sibling-calls -foptimize-strlen 
-fpartial-inlining -fno-peephole2 -fno-reorder-blocks-and-partition -fno-reorder-functions -frerun-cse-after-loop -fno-sched-interblock 
-fno-sched-spec -fno-schedule-insns -fno-strict-aliasing -fstrict-overflow -fno-tree-builtin-call-dce -fno-tree-switch-conversion 
-ftree-tail-merge -ftree-pre -fno-tree-vrp -fno-ipa-ra -freorder-blocks -fno-schedule-insns2 -fcode-hoisting -fstore-merging 
-freorder-blocks-algorithm=simple -fipa-bit-cp -fipa-vrp -fno-inline-functions -fno-unswitch-loops -fpredictive-commoning 
-fno-gcse-after-reload -fno-tree-loop-vectorize -ftree-loop-distribute-patterns -fno-tree-slp-vectorize -fvect-cost-model 
-ftree-partial-pre -fpeel-loops -fipa-cp-clone -fno-split-paths -ftree-vectorize --param early-inlining-insns=526 
--param gcse-cost-distance-ratio=12 --param iv-max-considered-uses=762
 -O3
--NCD:0.807842390787
---Test----
--Max:0
--Current:0
--Count:0
...

Results

The DB file saved in the PATH:/examples/gccflags/opentuner.db/Your PC's Name.db

Each sequence of compilation flags and the corresponding ncd value are saved in the db file.

Set up how many times to run

Please refer to the settings in main.py There are two strategies The default setting runs 100 times, if you want to modify it according to your own wishes this is ok. For example, by monitoring the change of NCD value in 100 times, if the cumulative change of 100 times increase is less than 5%, let's terminte it.

First-order formulas

We manually generate first-order formulas after understanding the compiler manual. The knowledge we learned is easy to move between the same compiler series---we only need to consider the different optimization options introduced by the new version.

We use Z3 Prover to analyze all generated optimization option sequences for conflicts and make changes to conflicting options for greater compiling success.

For more details, please refer Z3Prover.

Setting for Genetic Algorithm

The genetic algorithm is a metaheuristic inspired by the process of natural selection that belongs to the larger class of evolutionary algorithms. Genetic algorithms are commonly used to generate high-quality solutions to optimization and search problems by relying on biologically inspired operators such as mutation, crossover, and selection.

We tune four parameters for the genetic algorithm, including mutation_rate, crossover_rate, must_mutate_count, crossover_strength.

For more details, please refer globalGA.

Future Work

We are studying constructing custom optimization sequences that present the best tradeoffs between multiple objective functions (e.g., execution speed & NCD). To further reduce the total iterations of BinTuner, an exciting direction is to develop machine learning methods that correlate C language features with particular optimization options. In this way, we can predict program-specific optimization strategies that achieve the expected binary code differences.

BinTuner is a cost-efficient auto-tuning framework, which can deliver a near-optimal binary code that reveals much more differences than -Ox settings.

Related tags

Overview

BinTuner

The architecture of BinTuner:

System dependencies

Installation

Install Compiler

Checking Installation

Results

Set up how many times to run

First-order formulas

Setting for Genetic Algorithm

Future Work

Owner

BinTuner

Pandas-based utility to calculate weighted means, medians, distributions, standard deviations, and more.

scikit-survival is a Python module for survival analysis built on top of scikit-learn.

Find exposed data in Azure with this public blob scanner

Airflow ETL With EKS EFS Sagemaker

Average time per match by division

Sensitivity Analysis Library in Python (Numpy). Contains Sobol, Morris, Fractional Factorial and FAST methods.

We're Team Arson and we're using the power of predictive modeling to combat wildfires.

This creates a ohlc timeseries from downloaded CSV files from NSE India website and makes a SQLite database for your research.

Synthetic data need to preserve the statistical properties of real data in terms of their individual behavior and (inter-)dependences

Project: Netflix Data Analysis and Visualization with Python

INF42 - Topological Data Analysis

Pipeline to convert a haploid assembly into diploid

A set of functions and analysis classes for solvation structure analysis

InDels analysis of CRISPR lines by NGS amplicon sequencing technology for a multicopy gene family.

Port of dplyr and other related R packages in python, using pipda.

Analyzing Earth Observation (EO) data is complex and solutions often require custom tailored algorithms.

Evidence enables analysts to deliver a polished business intelligence system using SQL and markdown.

MS in Data Science capstone project. Studying attacks on autonomous vehicles.

Python tools for querying and manipulating BIDS datasets.

A distributed block-based data storage and compute engine