Home

Awesome

Krun

Krun is a framework for running high-quality software benchmarking experiments. Krun experiments consist of a configuration file, a carefully configured benchmarking machine, and the benchmarks themselves.

Step 1: Initial installation

Krun currently only runs on Debian Linux and OpenBSD. Porting it to other Unix variants is unlikely to be difficult, and we welcome patches.

Dependencies

You need to have the following programs installed:

If you choose to use pip for the Python dependencies, you can install them using pip -r requirements.txt.

If you want to benchmark Java, you will also need:

Step 2 (Linux only): kernel and OS setup

P-states

Benchmarking is at its most accurate when Intel p-states are disabled in the kernel. If you are using Grub, this can be achieved as follows:

If you are unable to do this, you can disable Krun's P-state check with --disable-pstate-check, but be aware that this degrades the quality of the resulting benchmarking numbers.

Performance counters

If you want low latency access to the following counters:

you will need to use the (now rather out of date) custom Linux kernel at:

https://github.com/softdevteam/krun-linux-kernel

If you do use this kernel you will need to set MSRS=1 in your Unix environment when building Krun (see later).

Tickless Kernel

Unless --no-tickless-check is passed to Krun, all cores but the boot core are expected to be in full adaptive ticks mode (tickless mode).

On older kernels which still have the CONFIG_NO_HZ_FULL_ALL config option, build the kernel with this set to "y".

On newer kernels, where CONFIG_NO_HZ_FULL_ALL is absent, you will have to add a nohz_full=1-N kernel command line option (where N is the number of cores your system has, minus one).

If your system has only one core then tickless mode is not checked, as the boot core (i.e. your only core) cannot be placed into adaptive ticks mode.

Step 3: Fetch and build Krun

First fetch Krun:

$ git clone --recursive https://github.com/softdevteam/krun.git
$ cd krun

Then run make (or gmake on OpenBSD). The Krun Makefile honours the standard variables: CC, CPPFLAGS, CFLAGS and LDFLAGS.

If you want to benchmark Java programs, you need to set the JAVA_HOME environment variable to point to your JDK installation, and set several other flags:

$ make clean
$ JAVA_HOME=/usr/lib/jvm/java-8-openjdk-amd64/ make  \
    JAVA_CPPFLAGS='"-I${JAVA_HOME}/include -I${JAVA_HOME}/include/linux"' \
    JAVA_LDFLAGS=-L${JAVA_HOME}/lib ENABLE_JAVA=1

Step 4: Audit system services

Background services (e.g. cron or sendmail) can interfere with benchmarking. The more services that you are able to switch off, the less interference is likely to occur. Some services are best disabled at boot and/or permanently (depending on your OS) and must be done manually. However, you may wish to disable some services only during benchmarking (e.g. you may wish to have a mail server running before and after benchmarking to inform you of benchmarking progress), which can be specified in the PRE_EXECUTION_CMDS and POST_EXECUTION_CMDS settings in your Krun config file.

Commands in each list are run, in order, using the krun user's shell (e.g. /bin/sh). If a command fails, Krun stops execution immediately without running subsequent commands. If you wish execution to continue even if a command fails you can use standard shell idioms: e.g. cmd || true guarantees that the overall command succeeds even if cmd fails.

For example on a systemd Linux you may turn daemons off before execution with:

PRE_EXECUTION_CMDS = [
    "sudo systemctl stop cron",
    "sudo systemctl stop atd",
    ...
]

and turn them back on with:

POST_EXECUTION_CMDS = [
    "sudo systemctl start cron || true",
    "sudo systemctl start atd || true",
    ...
]

In general it is best practise to turn things back on explicitly, because after Krun runs the final benchmark it will not reboot the machine. If, for example, you put network interfaces down in PRE_EXECUTION_CMDS, you should put them back up in POST_EXECUTION_CMDS so that you can login to the machine after the final benchmark has been run. We urge you to check such commands carefully: small oversights can easily lead to you locking yourself out of the system.

Krun can also copy intermediate results to a remote host and query that host to see whether it should suspend benchmarking. See https://github.com/softdevteam/warmup_experiment/blob/master/warmup.krun for more advanced options.

Linux

Note that Debian has, from a benchmarking perspective, the unfortunate habit of automatically starting daemons which get pulled in by dependencies.

On Linux, list services with:

# systemctl | grep running

Permanently disable services (including after system reset) with:

# systemctl stop <service>
# systemctl disable <service>

Commonly enabled services that you may wish to disable:

OpenBSD

On OpenBSD, list services with:

# doas rcctl ls started

Permanently disable services (including after system reset) with:

# rcctl stop <service>
# rcctl disable <service>

Commonly enabled services that you may wish to disable:

Note that Krun is unable to check whether turbo mode is enabled on OpenBSD or not, and is also unable to use APERF/MPERF ratios to indirectly check whether turbo mode was used. You should therefore be particularly careful to check that turbo mode is disabled when benchmarking on OpenBSD.

Step 5: Build and run the example

The examples directory contains the example.krun experiment. This contains two benchmark programs (nbody and dummy), both of which are run on C and PyPy. Each benchmark is run for 5 in-process iterations (where the benchmark is repeated 5 times within a for loop within a single process) across 2 process executions (where the entire VM is restarted).

First build the examples:

$ cd examples/benchmarks
$ pwd
.../krun/examples/benchmarks
$ make

If you also want to try the example Java benchmarks, you must build them as a separate step. Additionally, be sure to have compiled Krun with Java support as described in Step 3

$ pwd
.../krun/examples/benchmarks
$ make java-bench

Then run the example:

$ cd krun/examples
$ ../krun.py example.krun

You should see a log scroll past, and results will be stored in the file: ../krun/examples/example_results.json.bz2.

If you want to try the example Java benchmarks, there is a separate configuration file called java.krun, which contains configuration for the Java and Python examples:

$ ../krun.py java.krun

Note, this will only work if you have followed the earlier steps to compile Krun with Java support.

The Krun user

Krun runs benchmarks under a new Unix user krun, which is wiped and re-added before every experiment. Your Krun build and your experiment must both be readable by the krun user for the experiment to run.

Creating your own experiments

It is easiest to use examples/example.krun as a template for your own, new, experiments. Note that: the benchmarks referenced in the config file must be in a benchmarks subdirectory; and that each benchmark in the config must be in a subsubdirectory with a matching name. A typical directory structure is therefore as follows:

experiment/
    experiment.krun
    benchmarks/
        Makefile
        benchmark_1/
            language_1/
            benchmark_file.lang1
    ...

The top-level Makefile should build the VMs and benchmarks needed for the experiment.

Each benchmark should expose a function (or method) called run_iter which is the entry point to the benchmark. To preserve source code history, it can be easiest to put this function in a new file, which then imports the benchmark.

With regards to VM/compiler support, there are two ways Krun can invoke a benchmark:

The former option is best, as it supports Krun's core-cycle counting and APERF/MPERF ratio features. The following compilers/VMs are currently supported for this mode:

If your VM isn't listed, you can either add it to Krun, or use the external suite definition (see below). To add a new VM definition, add a new class to krun/vm_defs.py and a new iteration runner to the iterations_runners directory.

The latter option -- ExternalSuiteVMDef -- is useful if you want to quickly wrap an existing benchmark suite. For an example see examples/ext.krun and examples/ext_script.py.

To add a new platform definition, add a new class to krun/platform.py.

Testing your configurations

Before doing a full run of an experiment, you should perform a quick(ish) test of your configuration. This can be achieved with:

$ /path/to/krun/krun.py --dry-run --quick --debug=INFO config.krun

See the "Development and Debug Switches" section for a description of these switches.

Production benchmarking

Achieving the highest possible benchmarking quality requires more care. First, none of Krun's debug or development switches must be used. Second, Krun needs to run in "reboot mode" where each process execution will be run after the machine has (automatically) rebooted. The simplest way to do this is to have your init system invoke scripts/run_krun_at_boot via an rc.local script.

Your /etc/rc.local should look like this:

#!/bin/sh
/usr/bin/sudo -u someuser /path/to/krun/scripts/run_krun_at_boot /path/to/your/config.krun
exit 0

Make sure you replace the paths as appropriate and substitute someuser with the name of a normal unprivileged user that you wish to use to kick off Krun.

Make sure /etc/rc.local is executable and that it only contains absolute paths. Note that sudo(8) is installed in different places on different operating systems (for OpenBSD, it's /usr/local/bin/sudo).

Any arguments supplied after the config file path are passed to Krun unchanged.

You can then start the experiment by manually running the command from your new rc.local (i.e. sudo -u ...).

Monitoring progress

Whilst benchmarking is occuring, you must not log in to the machine (indeed, hopefully sshd, or equivalent, has been disabled!). To monitor progress, and be informed of errors, you you should add a MAIL_TO list of emails to your Krun config file:

MAIL_TO = ["me@mydomain.com", "other_person@herdomain.com"]

Krun uses sendmail(8) to send email, so you will need to make sure that you have a functional SMTP server installed (and don't forget to switch it off during benchmarking!).

Custom Dmesg Whitelists

For each platform, Krun has a default built-in dmesg whitelist. The whitelist is a collection of regex patterns which are used to decide if a line in the dmesg buffer is a cause for concern. If a new line appears in the dmesg during benchmarking, and the line is not matched by at least one whitelist pattern, then Krun will flag the process execution as ``errored'' and email the user.

From time to time you may find that you need to customise the whitelist. This is achieved by adding a callback named custom_dmesg_whitelist into your config file. The callback is passed the default list of patterns for your platform and must return a new list of patterns. In the implementation of your callback you have the choice to base your custom whitelist on the defaults or to define your own patterns from scratch.

For example, to add a pattern, you would add a callback like:

def custom_dmesg_whitelist(defaults):
    return defaults + ["^.*your+regex.*pattern$"]

Krun uses Python's re module to compile regex patterns. Consult the Python docs for more information on the regex syntax.

Bear in mind that Linux dmesg lines start with a time code which custom dmesg lines will need to match.

If you have added custom patterns which you think would be useful for other users of Krun, please raise an issue (or pull request) to have the patterns added to the defaults.

Development and Debug Switches

If you are making changes to Krun itself (for example, to add a new platform or virtual machine definition), there are a few switches which can make your life easier.

Krun results files

Krun generates a bzipped JSON file containing results of all process executions. The structure of the JSON results is as follows:

{
    'audit': '',  # A dict containing platform information
    'config': '', # A unicode object containing your Krun configuration
    'wallclock_times': {        # A dict containing timing results
        'bmark:VM:variant': [   # A list of lists of in-process iteration times
            [ ... ], ...        # One list per process execution
        ]
    },
    'core_cycle_counts': {      # Per-core core cycle counter deltas
        'bmark:VM:variant': [
            [                   # One list per process execution
                [...], ...      # One list per core
            ]
    },
    'aperf_counts': {...}       # Per-core APERF deltas
                                # (structure same as 'core_cycle_counts')
    'mperf_counts': {...}       # Per-core MPERF deltas
                                # (structure same as 'core_cycle_counts')
    'pexec_flags': {...}        # A flag for each process execution:
                                # 'C' completed OK.
                                # 'E' benchmark crashed.
                                # 'T' benchmark timed out.
    'eta_estimates': {u"bmark:VM:variant": [t_0, t_1, ...], ...} # A dict mapping
                  # benchmark keys to rough process execution times. Used internally:
                  # users can ignore this.
}

Some options exist to help inspect the results file:

$ python krun.py --dump-config examples/example_results.json.bz2
INFO:root:Krun starting...
[2015-11-02 14:23:31: INFO] Krun starting...
import os
from krun.vm_defs import (PythonVMDef, NativeVMDef)
from krun import EntryPoint

# Who to mail
MAIL_TO = []
...

$ python krun.py --dump-audit examples/example_results.json.bz2
{
    "cpuinfo":  "processor\t: 0\nvendor_id\t: GenuineIntel\ncpu family\t:
...

$ python krun.py --dump-reboots examples/example_results.json.bz2
[2015-11-06 13:14:35: INFO] Krun starting...
8

Troubleshooting

java.lang.UnsatisfiedLinkError Error

The following error in the log file is indicative that Krun has not been compiled with Java support:

Exception in thread "main" java.lang.UnsatisfiedLinkError: IterationsRunner.JNI_krun_init()V
	at IterationsRunner.JNI_krun_init(Native Method)
	at IterationsRunner.main(iterations_runner.java:248)

See Step 3 on how to build with Java support.

Unit Tests

Krun has a pytest suite which can be run by executing py.test in the top-level source directory.

Security Notes

Krun is not intended to be run on a secure multi-user system, as it uses sudo to elevate privileges. It also uses files with fixed names in /tmp/ which means that only one instance of Krun should be run at any one time (running more than one leads to undefined behaviour).

sudo is used to:

Please make sure you understand the implications of this.

Licenses

Krun is licensed under the UPL license.

The nbody benchmark is licensed under a revised BSD license: http://shootout.alioth.debian.org/license.php