-
Notifications
You must be signed in to change notification settings - Fork 74
Performance Tests
Authors: Hailey Pan, Zuozhi Wang
Reviewed by Chen Li
Set up benchmarks to test the performance of each operator in Texera. See new performance numbers after each pull request into the master.
As of 9/25/2016: FINISHED
Code in module: edu.ics.uci.texera.perftest
The packages dictionarymatcher, keywordmatcher, fuzzytokenmatcher, regexmatcher, and nlpextractor contain the performance test code of each operator.
The package runme contains the main function to start running the performance tests.
We are using the MEDLINE dataset. Its description and files are at here. The package medline contains the schema of the dataset.
Data files need to be put in the (texera directory)/texera/texera-perftest/sample-data-files folder. Put one data file in this folder; otherwise, it will affect how we display the results later.
The perftest-files/queries folder contains a file of sample queries, which is used in testing KeywordMatcher and DictionaryMatcher. The perftest-files/results folder contains the performance test results.
Write index and run performance tests: In the package runme,
-
WriteIndex.javawrites index. -
RunTests.javaassumes that an index already exists, and runs the performance tests. -
RunPerftests.javawrites an index first and then runs performance tests. The index is written into the (texera directory)/texera/texera-perftest/index folder.
As mentioned earlier, we want to automate the performance test process. So we write a Python script and use a cron job to run it automatically everyday. The python script build.py will pull changes from github, then run performance tests if there’s a change in the master branch.
It’s easy to run performance tests in an IDE (for example, Eclipse or IntelliJ). We can simply run the java file, and the IDE takes care of the rest. However, in a command line environment, it’s much harder to run the program. The command to run the program is generated in build.py. (Attention: the command needs to be changed if Texera's dependencies change.)
The Java performance test program writes the results into the "perftest-files/results" folder. There’s one csv file for each operator to record the results of each run.
Here’s a sample format of one csv file "keyword-phrase.csv":
| Date | Record # | Min Time | Max Time | Average Time | Std | Average Results | Commit Number |
|---|---|---|---|---|---|---|---|
| 09-09-2016 00:54:18 | abstract_100 | 0.017 | 1.373 | 0.2371 | 0.4464 | 2.18 |
Other operators’ csv files look similar to the format above.
The "Commit Number" column is empty because we choose to let the Python script fill in the commit number. So running the Java program, either via IDE or command line, won’t produce a commit number in the result file. The commit number is only added by running the Python script.
We use an open source Java package called "dashbuilder" to display the results. DashBuilder will automatically read the results produced by the python script, and display the results. We have an internal documentation file to describe how to set up the dashbuilder.