Skip to content

Latest commit

 

History

History
533 lines (441 loc) · 27.3 KB

CHANGELOG.md

File metadata and controls

533 lines (441 loc) · 27.3 KB
Raw

BenchExec Changelog

BenchExec 2.2

This release fixes two security issues, all users are encouraged to update:

  • Since BenchExec 2.1, the setup of the container for the tool-info module (which was added in BenchExec 1.20) could silently fail, for example if user namespaces are disabled on the system. In this case the tool-info module would be executed outside of the container. Run execution was not affected.
  • The kernel offers a keyring feature for storage of keys related to features like Kerberos and ecryptfs. Before Linux 5.2, there existed one keyring per user, and BenchExec did not prevent access from the tool inside the container to the kernel keyring of the user who started BenchExec. Now such accesses are forbidden (on all kernel versions) using seccomp if libseccomp2 is installed, which should the case on any standard distribution. Note that seccomp filters do have a slight performance impact and could prevent some binaries on exotic architectures from working. In such a case please file a bug report.

BenchExec 2.1

benchexec can now partition the Level 3 cache of the CPU for parallel runs and measure cache usage and memory bandwidth, at least on some Intel CPUs and if the pqos and pqos_wrapper are installed. More information is in the documentation.

Furthermore, some error messages for systems without container support were improved.

BenchExec 2.0

This release does not add new features compared to BenchExec 1.22, but removes several deprecated features and brings several other backwards-incompatible changes to make BenchExec more consistent and user-friendly:

  • Support for Python 3.2 and 3.3 is removed, the minimal Python version is now 3.4. Additionally, runexec/RunExecutor continue to support Python 2.7 until end of 2019.
  • Support for running benchmarks as a different user with sudo is removed (parameters --user/--users). Use container mode as better method for isolating runs.
  • Container mode is enabled by default. It can be disabled with --no-container, but this decreases reliability of benchmarking.
  • If the cpuacct cgroup is not available, CPU-time measurements and limits are not supported.
  • Either container mode or the freezer cgroup are required to ensure protection against fork bombs.
  • Niceness of benchmarked process is not changed, previously it was increased by 5.
  • Changes to input of benchexec:
    • The memory limit given to benchexec requires an explicitly specified unit.
    • Support for <test> tags, <sourcefiles> tags, and variables named ${sourcefile_*} removed from benchmark definitions. Use <rundefinition>, <tasks>, and ${inputfile_*} instead.
    • Variables named ${taskdef_*} are defined only if task-definition files are used, and variables named ${inputfile_*} only otherwise.
  • Changes to table-generator:
    • A column named memUsage is automatically renamed to memory.
    • A column named memory is automatically converted to Megabytes. Both conversions are only applied if no <column> tags are used.
  • Changes to run-result data:
    • In case of aborted or failed runs, no dummy results (e.g., cputime of 0s) are present.
    • The memory results of benchexec are named memory, not memUsage.
    • Memory results have the unit B explicitly specified. Furthermore, units are present in all attributes of the result XML files where they were still missing.
    • Result item exitcode is removed, only RunExecutor.execute_run() still returns it, but as an object instance instead of an int. Use returnvalue and exitsignal instead.
  • Module benchexec.test_tool_wrapper is removed, use benchexec.test_tool_info instead.
  • BenchExec (both benchexec and runexec) terminates itself cleanly after aborting all runs if it receives one of the signals SIGTERM, SIGINT (Ctrl+C), or SIGQUIT.

Additionally, this release adds a fix for the container that is used since BenchExec 1.20 for the tool-info module. In this container, the environment variable HOME did not point to /home/benchexec as expected but to the user's real home directory. This broke tools like Ultimate if the /home was configured to be hidden or read-only.

Furthermore, we declare the following features deprecated and plan on removing them for BenchExec 3.0, which is expected to be released in January 2020:

  • Support for Python 2.7 and 3.4 (cf. #438)
  • Support for checking correctness of run results and computing scores if task-definition files are not used (cf. #439)

Please respond in the respective issue if one of these deprecations is a problem for you.

BenchExec 1.22

  • More robust handling of Ctrl+C in benchexec. For example, output files are now always fully written, whereas previously pressing Ctrl+C at the wrong time could result in truncated files. A side effect of this is that if you call benchexec.benchexec.BenchExec().start() in own Python code, you must now add a signal handler for SIGINT. The same was already true for users of RunExecutor, this is now documented.
  • Fix Ctrl+C for benchexec in container mode. In BenchExec 1.21, one would need to press Ctrl+C twice to stop benchexec.
  • Fix unreliable container mode on Python 3.7.
  • Some robustness improvements and fixes of rare deadlocks.
  • Decreased overhead of benchexec while runs are executing.

BenchExec 1.21

This release contains only a few bug fixes:

  • Forwarding signals to the benchmarked process (and thus, stopping runs via Ctrl+C), was broken on Python 2.
  • If the freezer cgroup was available but mounted in a separate hierarchy, it was not used reliably as protection against fork bombs when killing processes.
  • Since BenchExec 1.19, an exception would occur if a non-existing command was started in container mode.
  • Since BenchExec 1.19, copying output files from a container would occur while subprocesses are still running and would be counted towards the walltime limit. This is fixed, although subprocesses will still be running if the freezer cgroup is not available (cf. #433).

BenchExec 1.20

  • If benchexec --container is used, all code that is part of the tool-info module (as well as all processes started by it) are now run in a separate container with the same layout and restrictions as the run container. Note, however, that it is not the same container, so any modifications made by the tool-info module to files on disk are not visible in the runs! The test_tool_info utility also has gained a parameter --container for testing how a tool-info module behaves in a container.
  • Nested containers are now supported. Due to a change to the internal implementation of the container mode, commands like the following succeed now: containerexec -- containerexec --hidden-dir /sys -- /bin/bash. (Some parts of /sys need to be excluded because of kernel limitations.) Note that nesting runexec or benchexec is still not supported, because nested cgroups are not implemented, so any cgroup-related features (resource limitations and measurements) are missing. But nesting containerexec and runexec --container (or vice-versa) now works.
  • /etc/hostname in container now also shows the container's host name that exists since BenchExec 1.19.
  • Change how CPUs with several NUMA nodes per CPU are handled: BenchExec will now treat each NUMA node like a separate CPU package and avoid creating runs that span several NUMA nodes. Thanks @alohamora!

BenchExec 1.19

  • In container mode, all temp directories are now on a tmpfs "RAM disk". This affects everything written to directories in the hidden or overlay modes. Files written there are now included in the memory measurements and the memory limit! The advantage is that performance should be more deterministic, especially if several runs use much I/O in parallel. This feature can be disabled with --no-tmpfs.
  • /dev/shm and /run/shm are now available inside the container and provide a tmpfs instance (even with --no-tmpfs) as required by some tools for shared memory.
  • Container mode now recommends LXCFS and automatically uses it if available for a better container isolation (e.g., uptime measures container uptime, not host uptime). On Debian/Ubuntu, just use sudo apt install lxcfs.
  • Several small bug fixes and other improvements of isolation for container mode (e.g., host name in container is no longer the real host name).
  • Add benchexec --no-hyperthreading, which restricts core assignments to a single virtual core per physical CPU core (all other sibling cores will stay unused). Thanks @alohamora!

BenchExec 1.18

  • Add result done that tools can output if the standard results true/false/unknown are not applicable (for example because no property was checked), and the run completed successfully.
  • In container mode, --keep-system-config is no longer necessary if overlayfs is not used for /etc, and thus it is is no longer automatically implied in such cases.
  • Benchmark definitions support a new attribute displayName with a human-readable name that will be shown in tables.
  • A new variable ${taskdef_name} can now be used in places where variable substitution is supported.
  • Table-generator supports % as unit for numerical values.
  • Some improvements for score handling outside of SV-COMP (i.e., if scores are not calculated by BenchExec).
  • New tool-info modules for Test-Comp'19
  • Several small bug fixes and improvements

BenchExec 1.17

  • Tasks can now be defined in a YAML-based format, cf. the documentation This supports tasks with several input files, and allows providing metadata such as expected verdicts in a structured format instead of encoded in the file name. The format will be extended to handle more information in the future.
  • The wall-time limit can now be specified separately from the CPU-time limit for benchexec as command-line parameter or in the benchmark definition.
  • Support for SV-COMP'19 property memcleanup.
  • In containers, properly handle /run/systemd/resolve, which is necessary for DNS resolution on systems with systemd-resolved.
  • Avoid warnings for mountpoints below inaccessible directories in containers.
  • Improvements for handling NaN and Inf values in table-generator.
  • Log output of BenchExec will now have colors if coloredlogs is installed.
  • New tool-info modules and updates for SV-COMP'19.

BenchExec 1.16

BenchExec 1.15 (skipped)

BenchExec 1.14

  • Updated tool-info modules for all participants of SV-COMP'18.
  • Extended support for variable replacements in table-definitions of table-generator.

BenchExec 1.13

  • For Debian/Ubuntu, the .deb package is now the recommended way of installation, because it automatically configures cgroups as necessary.
  • BenchExec now automatically attempts to use the sub-cgroup system.slice/benchexec-cgroup.service if it does not have access to the current cgroup. This means that if you followed our installation instructions for systems with systemd, there is no need anymore to manually put your shell into the correct cgroup.
  • Several smaller bug fixes for table-generator: #249, #250, #259, #260, #271, #272
  • For users of the Python API of RunExecutor, different file names can now be specified for stdout and stderr of the tool.
  • Some new tool-info modules and updates for SV-COMP'18.

BenchExec 1.12

  • Fix execution of runs specified with <withoutfile> tags in the benchmark definition: the name of the run was missing from the command-line in BenchExec 1.11.

BenchExec 1.11

  • table-generator can now be given result XML files as arguments in addition to a table-definition XML file (with parameter -x). In this case, it will use the column definitions from the latter for tables with the separately given results.
  • The directory contrib of the repository now contains a script statistics-tex.py, which can export summary data for benchmark results (e.g., number of solved tasks, average CPU time, etc.) to LaTeX.
  • The dummy tools true and false, which could be used for testing a BenchExec installation, are replaced with a more generic dummy tool called dummy.
  • A few minor bug fixes and performance optimizations.

A new paper about BenchExec called Reliable Benchmarking: Requirements and Solutions is now available.

Please note that support for Python 3.2 and 3.3 is deprecated. Furthermore, the support for "sudo mode" (parameter --user/--users) is also deprecated. All deprecated features will be removed in BenchExec 2.0.

BenchExec 1.10

This release brings several smaller and medium-sized features:

  • Tool-info modules for all participants of SV-COMP'17, and support for results of the category correct-unconfirmed, which is used by SV-COMP if witness validation was not successful. To conform with SV-COMP's definitions, violations of the SV-COMP reachability property unreach-call will now be reported as false(unreach-call) instead of false(reach).
  • Measurement of block I/O if the blkio cgroup controller is available (experimental, please read the documentation!).
  • Measurement of the energy used by the CPU for a run, if the tool cpu-energy-meter is installed on the system (experimental, please read the documentation!).
  • Ability to limit the disk space a tool can occupy in container mode.
  • Various minor improvements to make container mode more robust.
  • The feature for executing benchmarks under different user accounts with sudo is now marked as deprecated and may be removed in the future, consider using the container mode instead for isolating runs (cf. issue #215).
  • table-generator is now more flexible:
    • Builtin support for certain unit conversions, such that the scale factor does not always need to be explicitly specified. Furthermore, unit conversions now work even if the values already have a unit.
    • Column titles can be manually specified with the displayTitle attribute.
    • What columns are relevant for the "diff" table can be configured.

Please also note that we are considering dropping the support for Python 3.2 and maybe 3.3 in BenchExec 2.0 (to be released in a few weeks). If this is a problem for you, please tell us in issue #207.

BenchExec 1.9

The main feature of this release is the addition of a container mode that allows to isolate runs from each other and from the host, for example preventing filesystem and network accesses. It also allows to collect and store all files created by the tool in a run. The container mode is still in beta and disabled by default for now, it will be enabled by default in BenchExec 2.0. Please try it out and tell us your experiences!

Further changes:

  • table-generator now supports HTTP(S) URLs to be given for result XML files to allow generating tables for results without needing to download them first. The HTML tables will contain correct links to the log files.
  • New SV-COMP property deadlock supported by benchexec.
  • The parameters --rundefinition and --tasks of benchexec now support wildcards.
  • Rounding of very small and very large values in table-generator has been fixed.
  • The default font for HTML tables has changed, it is now a font that supports correctly aligned digits.

BenchExec 1.8

  • benchexec now compresses results by default: XML result files are compressed with BZip2, and log files are stored within a ZIP archive. This can reduce the necessary disk space significantly (typically these logs compress very well), and for large benchmark sets it reduces the number of necessary files, which can make dealing with the results much faster. The previous behavior can be restored with the parameter --no-compress-results.
  • table-generator now supports benchmark results where the log files are stored in a ZIP file instead of a regular directory. All features continue to work with compressed results, including extraction of values from log files and viewing log files from HTML tables (cf. table-generator documentation for more details). Compressed and uncompressed results are handled transparently and can be mixed, and using results that were manually compressed or decompressed is also supported.

BenchExec 1.7

  • Fix table-generator behavior for columns where different cells have different units: The release notes for 1.6 claimed that these columns are treated as text column, when instead they were rejected. Now they are treated as text. Note that BenchExec does not create such columns itself, so this should not affect most users.
  • Fix computation of scores according to the SV-COMP scoring scheme: if the expected result is for example false(valid-deref) and the tool returns false(valid-free), the resulting score is the one for a wrong false answer (-16 points), not the one for a wrong true answer (-32 points). The latter score is only given if the tool actually answers true incorrectly.
  • Change result classification, if the returned answer does not belong to the property of the task, for example, if the tool returns true instead of sat for a task with category satisfiability, or if the tool returns false(no-overflow) when it should not even check for overflows. Now these results are classified as unknown (with score 0), previously these were treated as wrong answers.
  • Fix escaping of links in HTML tables, e.g., to log files with special characters in their name. This was broken in 1.6.

BenchExec 1.6

This release brings several improvements to table-generator:

  • table-generator now rounds measurement values in a scientifically correct way, i.e., with a fixed number of significant digits, not with a fixed number of decimal places. The attribute numberOfDigits of <column> tags in table-definition files now also specifies significant digits, not decimal places. By default, in HTML tables all fractional values are now rounded (e.g., time measurements) and all integer values continue without rounding (e.g., memory measurements), previously only "time" columns were rounded. The remaining rounding-related behavior stays unchanged: In CSV tables, values are not rounded by default, and if numberOfDigits is explicitly given for a column, it's value will always be rounded in both HTML and CSV tables.
  • table-generator now automatically extracts units from the cells in a column and puts them into the table header.
  • In HTML tables, numeric values are now aligned at the decimal point, and text values are left aligned (previously both were right aligned).
  • table-generator now allows to convert values from one unit into another. So far this is only implemented for values that do not have a unit attached to them, and both the target unit and the scale factor need to be specified explicitly in the <column> tag. This can be used for example to show memory measurements in MB instead of Bytes in tables.
  • table-generator now allows columns with links to arbitrary files to be added to tables.
  • table-generator does not handle columns where cells have differing units wrongly anymore. Previously, the unit was simply dropped, leading to wrong values for statistics. Now such columns are treated as text and no statistics are generated. (Note that BenchExec never creates such columns by itself, only if values are extracted from the tool output this could happen).

Other changes:

  • The behavior of benchexec --timelimit was changed slightly, if a value for hardtimelimit was given in the benchmark-definition file. If a time limit is specified on the command line, this now overrides both soft and hard time limit.
  • Implementation of tool-info modules got easier because the test_tool_info helper got improved (it now allows to test the function for extracting results from tool outputs).
  • Several tool-info modules of tools participating in SV-COMP got improved.
  • Simplified cgroups setup for systemd systems.
  • Improved documentation.

BenchExec 1.5

  • Improved definition of time and memory limits: Both can now be specified including units such as "s", "min" / "MB", "GB". to make them easier to read and less ambiguous. The old input format without units is still valid.
  • runexec now allows enabling other cgroup subsystems and setting arbitrary cgroup options.
  • HTML tables gained the possibility for inverting row filters.
  • Improve detection of out-of-memory situations (were not always reported as OOM).
  • External resources in HTML tables are loaded from HTTPS URLs such that browsers do not complain because of mixed content when viewing tables via HTTPS.
  • Improved warnings for swapping and CPU throttling for benchexec.
  • Various improvements to internal handling of memory values, they are not consistently stored as bytes (this only affects extensions of BenchExec, not regular input and output for users).

BenchExec 1.4

  • BenchExec moved to https://github.com/sosy-lab/benchexec
  • Fix several bugs in table-generator introduced in version 1.3.
  • BenchExec now creates fresh empty directories for $HOME and $TMPDIR of all runs, and removes them afterwards.
  • table-generator now transparently supports result XML files as input that are compressed with GZip or BZip2.
  • benchexec now reports some more information as status when a tool crashes, e.g. whether it segfaulted or aborted, and what the exit code was (previously this was only done for some tools).
  • If a tool produces a result but still violates a resource limit, this is now shown in the status (but still counted as timeout / out of memory).
  • Added dummy tool "calculatepi" that needs no input files and no installation, but can be used to create some CPU load and test benchmarking (it calculates Pi up some arbitrary number of digits using the tool "bc").
  • Renaming "tool wrapper" to "tool info". This is mostly an internal and documentation change, but the utility benchexec.test_tool_wrapper is now named benchexec.test_tool_info.

BenchExec 1.3

  • Fix core assignment on AMD Bulldozer/Piledriver Opterons.
  • Measure and report CPU time usage per core (hidden by default in tables, use table-generator --all-columns to show).
  • Parameter --user allows executing benchmarks under a different user (cf. https://github.com/sosy-lab/benchexec/blob/master/doc/separate-user.md).
  • Performance improvements for table-generator, including parallel processing of input and output files and statistics.
  • HTML Tables support filtering rows by task name.
  • Improved statistics in HTML tables: median is now the arithmetic median, unnecessary rounding removed, standard deviation added, and missing results are not counted as "0" but ignored in calculation.
  • New utility for testing tool wrappers, making it easier to add support for new tools.
  • Several new modules for integration of various software verifiers.

BenchExec 1.2

  • BenchExec now records whether TurboBoost was enabled during benchmarking.
  • Updated SV-COMP scoring scheme to SV-COMP 2016.
  • Support new property 'no-overflow' for SV-COMP 2016.
  • Several new modules for integration of various software verifiers.
  • Some improvements to CPU-core assignment.

BenchExec 1.1

  • HTML tables produced by table-generator now have a header that stays always visible, even when scrolling through the table.
  • A Debian package is now created for releases and made available on GitHub.
  • Small bug fixes.

BenchExec 1.0

  • Multiple runs for the same file can now be shown in the table in different rows if they have different properties or ids.
  • Helper files for generating scatter and quantile plots with Gnuplot added.
  • Doctype declarations are now used in all XML files.
  • Statistics output at end of benchexec run was wrong.

BenchExec 0.5

  • Allow to redirect stdin of the benchmarked tool in runexec / RunExecutor
  • Fix bug in measurement of CPU time (only occurred in special cases and produced a wrong value below 0.5s)
  • Improve utility command for checking cgroups to work around a problem with cgrulesngd not handlings threads correctly.

BenchExec 0.4

  • Support for integrating SMTLib 2 compliant SMT solvers and checking the expected output.
  • runexec now supports Python 2 again.
  • table-generator allows to selected desired output formats and supports output to stdout.
  • Added utility command for checking if cgroups have been set up correctly.
  • Avoid "false posititive/negative" and use "incorrect false/true" instead.
  • Command-line arguments to all tools can be read from a file given with prefix "@".
  • Bug fixes and performance improvements.

BenchExec 0.3

  • HTML tables now have header with direct access to plots.
  • Maximum score of table is generated again.
  • table-generator can now extract statistic values for other tools, too (not only CPAchecker).
  • More flexible time limit specifications.
  • Warnings shown if system swaps or throttles during benchmarking.
  • Improved reliability of benchmarking: forbid swapping, use freezer to kill processes atomically.
  • Renamed <sourcefiles> tag to <tasks> in benchexec input.
  • Bug fixes.
  • Added documentation.
  • Added more tests.

BenchExec 0.2

  • bug fixes
  • switch to Python 3 completely

BenchExec 0.1

Initial version of BenchExec as taken from the repository of CPAchecker.