CPROVER Benchmarking Toolkit
- Patch set management tools
- Benchmark execution helpers
- Result evaluation
Please report problems, bugs, questions, suggestions to Michael Tautschnig.
Download our benchmarking framework as tarball or as Debian/Ubuntu package. When using the .tar.gz archive, extract the archive and add cpbm-0.5 to your PATH enviroment variable. For the Debian/Ubuntu package: install it using dpkg -i cpbm_0.5-1_all.deb.
Benchmarks consist of two files: (1) an original source archive as obtained from its original authors (we provide local backup copies below), without modification and (2) a corresponding patch set <NAME>.cprover-bm.tar.gz that contains all modifications, execution scripts, and other helpers.
|Benchmark||Original Source||Patch Set|
|Conflict Driven Learning||Leo's page
|Loopfrog: Lincoln Labs Corpora||Lincoln Labs Corpora
|Loopfrog: Verisec||Verisec Suite
|NEC Labs: Atomicity Violations (1)||Chao Wang
|NEC Labs: Atomicity Violations (2)||Chao Wang
|LLBMC VSTTE'12 Benchmarks||llbmc.org
Benchmark Execution Example 1: DDVerify
Download the patch set.
Download and unpack the kernel sources and apply the patch set
$ cpbm unpack \ http://www.kernel.org/pub/linux/kernel/v2.6/linux-2.6.19.tar.bz2 \ linux-2.6.19-ddverify.cprover-bm.tar.gz
Enter the new directory
$ cd linux-2.6.19-ddverify
Run the default benchmark configuration and produce a LaTeX table
$ cprover/rules -j4 table
Produce an additional GNUplot script and display the graph
$ cprover/rules graph | gnuplot
Build a web page providing an overview table and all log files in cprover/results.satabs.dfl.web/:
$ cprover/rules web
Benchmark Execution Example 2
Take a look at the complete workflow using examples from the SAMATE Reference Dataset.
This ScreenFlow video requires a more recent version of the Adobe Flash Player to display. Please update your version of the Adobe Flash Player.
Patch Set Management
As described above, benchmarks consist of
(1) an original source archive, without modification and
(2) a corresponding patch set <NAME>.cprover-bm.tar.gz that contains all
modifications, execution scripts, and other helpers.
This design is inspired by Debian v3 source packages and may thus sound familiar to people who are used to working with Debian packages.
(1) The original source archive must be a .zip, .tar.gz, .tgz, .tar.bz2, or .tar archive. If the authors of the software to be benchmarked do not offer such an archive, it should be built manually from whatever the authors provide as source.
(2) The patch set <NAME>.cprover-bm.tar.gz is created and managed by cpbm update as described below. The archive ships a cprover/ directory that may contain arbitrary files, but at least a Makefile cprover/rules must be provided. The patch set management scripts will populate a cprover/patches directory that precisely describes all modification to original sources.
The basic design of this archive solution is completely independent of the software to be benchmarked. Its main purpose is the precise description of changes to the original source that were made in order to benchmark some tool.
Patch set management commands
cpbm unpack: Takes an archive of original source (or a URL to download the archive from) and the corresponding patch set <NAME>.cprover-bm.tar.gz. cpbm unpack then builds the directory <NAME> from the contents of the original source archive, unpacks the cprover directory in
, and applies patches from cprover/patches, if any.
cpbm init and cpbm update: Create and maintain a <NAME>.cprover-bm.tar.gz file. This file is initially created in the parent directory by running cpbm init <SOURCE ARCHIVE> from within directory <NAME> resulting from manually unpacking <SOURCE ARCHIVE>. This step creates the cprover/ directory and populates it with a template cprover/rules file.
Once the patch set <NAME>.cprover-bm.tar.gz has been created for the first time, it will only be updated by cpbm update. To this end, cpbm update <SOURCE ARCHIVE> unpacks the source archive into a temporary directory, applies patches previously recorded in cprover/patches, and computes the diff between the current working directory and the contents of the patched source.
If new changes are found, these are recorded as new patches in cprover/patches. The sequence of patches to be applied is stored in cprover/patches series. It is recommended that automatically created patches are renamed to more descriptive names. Consequently such renamings must be reflected in cprover/patches/series.
Creating a new benchmark suite, e.g., for the Linux kernel:
Download the original sources from kernel.org
Manually unpack them:
$ tar xjf linux-2.6.19.tar.bz2
We rename the directory to get proper benchmark name "linux-2.6.19-foo"
$ mv linux-2.6.19 linux-2.6.19-foo $ cd linux-2.6.19-foo
Create the basic patch set linux-2.6.19-foo.cprover-bm.tar.gz
$ cpbm init ../linux-2.6.19.tar.bz2
Edit some source files ... and record patches
$ cpbm update ../linux-2.6.19.tar.bz2
To make the benchmark source and all patches available for others to use publish linux-2.6.19-foo.cprover-bm.tar.gz and the original source archive or its URL.Using the benchmark suite:
$ cpbm unpack \ http://www.kernel.org/pub/linux/kernel/v2.6/linux-2.6.19.tar.bz2 \ linux-2.6.19-foo.cprover-bm.tar.gz $ cd linux-2.6.19-foo
cpbm init, as described above, creates as template the cprover/rules file (a
Makefile). For most basic use cases filling in the names of the benchmarks to be
run (e.g., all C files without the .c suffix) in the BENCHMARKS = XXX line and
choosing a suitable default configuration in CONFIG = XXX (see examples in there)
will suffice to produce a working benchmark package.
The file may, however, be fully customized as deemed necessary. The only assumption made by cpbm is that cprover/rules clean performs a cleanup. Remember to run cpbm update before distributing the .cprover-bm.tar.gz file.
The actual benchmark execution is then triggered by
$ cprover/rules -j4 verifyto perform verification of all benchmark instances in 4 parallel threads of execution with the default configuration. This step first induces a build of each benchmark from source, if necessary.
To choose a different configuration for the verification run, override the CONFIG variable. For instance, to perform verification using CPAchecker's explicit analysis use
$ cprover/rules -j4 verify CONFIG=cpachecker.explicit
Benchmark Execution Helpers
Benchmark execution makes use of a number of helpers:
cpbm list-claims: the CPROVER tools are able to selectively verify a chosen claim in a program under scrutiny. This tool lists the possible claims as pairs <CLAIM>:<STATUS>, where <CLAIM> is the identifier used by the verification tool and <STATUS> is either UNKNOWN, or TRUE in case a claim is trivially satisfied.
If the verification CONFIG does not support specific claims (as is the case for all non-CPROVER tools), ALL_CLAIMS:UNKNOWN is return as pseudo claim identifier and verification status.
cpbm run: actually executes the verification tool with the configured options and produces a log file. This log file contains the output of the verification tool plus environment information and statistics. All further benchmark evaluation is based on such log files.
The names of all log files produced by cpbm run will be listed in cprover/verified.$CONFIG. Performing
$ cprover/rules csv
yields a CSV (comma-separated value) formatted summary of all results in cprover/results.$CONFIG.csv. This file may be read using most spreadsheet programs for further manual inspection and evaluation. With cpbm, however, also LaTeX tables, GNUplot graphs, or web pages may be produced:
$ cprover/rules table
yields a LaTeX summary of execution times and memory usage of all benchmark instances. With
$ cprover/rules graph
furthermore LaTeX/TikZ code is produced which may be included in publications or taken verbatim to produce PDF files.
$ cprover/rules web
collects all log files in a new directory cprover/results.$CONFIG.web plus an index.html file that contains an HTML formatted version of the CSV table. Each benchmark links to the respective log file.
These steps permits a number of customizations:
The CSV data is produced by cpbm csv, which uses a specific parser of the output for each verification tool. These parsers are perl scriptlets found in the parse-*.pl file of the CPROVER benchmarking toolkit distribution.
To produce additional columns in the CSV file, add further patterns to these parsers or copy and adapt one of the existing parsers. New parsers are preferably added to the distribution, but can also be put in the cprover/ directory. If a parser exists both in cprover/ and in the distribution, the former will be used.
The LaTeX table is generated using cpbm table, which takes as arguments a CSV file and one or more column names that shall be included (in the specified order) in the LaTeX output. Consequently, adding further columns to CSV output as described above also permits printing this output to LaTeX.
The graphs may either be box plots of CPU time and memory usage or scatter plots (using cpbm graph -s) for comparison of two tools. For box plots an arbitrary number of CSV files may be specified, whereas scatter plots require exactly two CSV files.