|
…
|
||
|---|---|---|
| .. | ||
| artifactdownloader | ||
| mach_perftest_notebook_dev | ||
| Dockerfile | ||
| README.md | ||
| __init__.py | ||
| ptnb-config-linux-new.yml | ||
| run_variance_analysis.py | ||
| variance_analysis.py | ||
README.md
moz-current-tests - variance analysis
This subfolder contains a tool that can be used to analyze the variance between two CI test runs.
Setup
The first time you run, you'll need to run
cd variance-analysis/mach_perftest_notebook_dev
python3 setup.py develop
cd ..
Running a Variance Analysis
Here's a sample command that can be used:
python3 run_variance_analysis.py --base-revision 5e4047061e46c5cb86d1ef694bc206fc8f4e7d20 --new-revision 5e4047061e46c5cb86d1ef694bc206fc8f4e7d20 --base-branch autoland --new-branch autoland --tests cnn --platform linux --output testing --config ptnb-config-linux-new.yml --search-crons
Only one platform may be specified at a time, but multiple tests are ok. The tests can be specified as exact matches or a substring of the test (e.g. cnn). The config can be used to specify the Transformer to use on the data - all of fields get replaced. Be very careful with the test and platform selection to ensure that you aren't mixing data. In the majority of cases, the ReplicateTransformer should take care of this by combining the Taskcluster task name with the extra options found in the perfherder-data. The number of artifacts downloaded will tell you if you are downloading the right things (1 artifact per task).
The base revision should be a CI run without changes and the new revision should point to the CI run with changes.
Run python3 run_variance_analysis.py --help for more help and options.
At the end of the run, you will see a list of the subtests that were analyzed and you will also see a table like this output with information about the variance:
linux - warm total: X/18
Average noise diff: ...
Significance of the average: ...
Number of tests with lower noise: ...
Number of tests with higher noise: ...
Averager decrease in noise for lower: ...
Average increase in noise for higher: ...
The Average noise diff gives a ratio between the new and the base grouping. A ratio greater than 1 implies that the noise is larger in new
linux - cold total: Y/18
Average noise diff: ...
Significance of the average:...
Number of tests with lower noise: ...
Number of tests with higher noise: ...
Averager decrease in noise for lower: ...
Average increase in noise for higher: ...
The Average noise diff gives a ratio between the new and the base grouping. A ratio greater than 1 implies that the noise is larger in new