To start benchmarking, try:
python benchmark.py [project] -[options]
Results will be kept in a similarly-named database, e.g. [project].db
Commands and their output will be kept in a similarly-named log, e.g logs/[project].log
Generally, create a benchmarks
directory in your project's repository. Within that, create files that start with benchmark_
, e.g. benchmark_item.py
. Follow the general rules for a unit test, and create classes with functions that start with the prefix benchmark_
. If these naming conventions are followed, benchmarks will be found and run automatically.
[project].json
will be read in if command python benchmark.py [project]
is run and [project].json
exists. If [project].json
does not exist, benchmark will fail. Here are the JSON fields currently supported in that file:
repository
: the repo to be benchmarked.
branch
: the branch of that repo to benchmark.
triggers
: list of repositories, that if changed, should trigger a new benchmarking run.
dependencies
: things needed by the repository
in the conda env to successfully run the benchmark.
An example [project].json
file:
{
"repository": "[email protected]:user/repository",
"branch": "work",
"triggers": [
"[email protected]:user/trigger_repo_1",
"[email protected]:user/trigger_repo_2",
"https://bitbucket.org/user/trigger_repo_3"
],
"dependencies": [
"python=2.7",
"numpy",
"scipy",
"networkx",
"sqlitedict",
"mpi4py",
"git+https://bitbucket.org/petsc/[email protected]",
"git+https://bitbucket.org/petsc/[email protected]"
]
}
--plot
, -p
: To plot a specific spec from the database, after benchmarking has been run, use python benchmark.py [project] --plot [spec]
--keep-env
, -k
: To keep your temporary conda environment around after benchmarking, usually for troubleshooting purposes. Note: when your run ends, the env will be kept, but you'll be returned back to whatever env you started the run in. To inspect the kept env, do a conda env list
and source activate [envname]
--force
, -f
: To force a run of the benchmarks even if no trigger repos have been changed, simply add --force
to the command. This is usually used for testing, or if something went wrong with a previous run and you want the benchmarks to run again.
--unit-tests
, -u
: Runs the unit tests before running the benchmarks. Benchmarks running are contingent on unit tests passing.
--dump
, -d
: Dump the contents of the database to an SQL file.
benchmark.cfg
will be used if it exists, but is generally only needed if administering a benchmark "server."
JSON file currently used to specify fields such as:
env
: set certain environment variables before use.
slack
: post benchmarking results message to channel using token. (see JSON example below and Slack documentation for more details on custom integrations/hooks.)
images
: can be used to upload images to a specific url.
data
: can be used to specify a place to upload a database backup.
working_dir
: where to look for everything. defaults to where benchmark.py lives.
repo_dir
: where to clone the repositories of the main and trigger repos. Defaults to working_dir/repos
logs_dir
: where to keep the logs for all of the projects being benchmarked. Defaults to working_dir/logs
remove_csv
: remove benchmark data file after adding to database. Default: False
.
plot_history
: generate a plot showing history of each benchmark. Default: True
.
ca
: CA cert information needed by curl, with two sub-data:
cacert
: Defaults to "/etc/ssl/certs/ca-certificates.crt"
capath
: Defaults to "/etc/ssl/certs"
url
: the base URL for linking to a benchmark website/URL handler running tornado_server.py
An example benchmark.cfg
:
{
"env": {
"LD_PRELOAD": "/path/to/libgfortran.so.3"
},
"slack": {
"message": "https://hooks.slack.com/services/[hook]/[address]",
"file": "https://hooks.slack.com/services/[hook]/[address]",
"channel": "#benchmarks",
"token": "slack-token-serial-number"
},
"images": {
"url": "http://domain.com/path/to/images",
"upload": "[email protected]:directory/subdir"
},
"data": {
"upload": "[email protected]:directory"
}
"url" : "http://domain.com/benchmark/"
}
It's suggested that if you want to run benchmark
on a regular basis to keep track of history of code changes and how they change your performance, that you set benchmark
to run on a regular schedule, for instance, in a cron
job.
Given some of the underlying tools used to make benchmark work, it will likely only work with Linux and MacOSX.