NOTE: This is still a work in progress project, suitable for:
developers, contributors and testers.
None of the provided tests should be considered stable and/or suitable
for the evaluation of a product.
The LISA project provides a toolkit that supports regression testing and interactive analysis of workload behavior. LISA stands for Linux Integrated/Interactive System Analysis. LISA's goal is to help Linux kernel developers to measure the impact of modifications in core parts of the kernel. The focus is on the scheduler, power management and thermal frameworks. However LISA is generic and can be used for other purposes too.
LISA provides an API for modeling use-cases of interest and developing regression tests for use-cases. A ready made set of test-cases to support regression testing of core kernel features is provided. In addition, LISA uses the excellent IPython notebook framework and a set of notebooks are provided for live experiments on a target platform.
This is an overall view of the toolkit:
+---------------------------------------+
| +-----------------------------------+ |
| | LISA Toolkit | |
| | +-----------------------------+ | |
| | | IPython Notebooks | | |
| | +-----------------------------+ | |
| | +-------+ +-------+ +---------+ | |
| | | wlgen | | tests | | reports | | |
| | +-------+ +-------+ +---------+ | |
| +-----------------------------------+ |
| +----------+ +-------+ |
| | | | BART | |
| | | +-------+ |
| | devlib | +----------------------+ |
| | | | TRAPpy | |
| +--^-------+ +----------------------+ |
| || |
| || HOST |
+---------------------------------------+
||
|| SSH/ADB
|+
+---V-----------------------------------+
| TARGET |
| |
| Linux/Android |
| or Localhost |
+---------------------------------------+
LISA depends on the following external python libraries:
This note assumes installation from scratch on a freshly installed Debian system.
$ sudo apt-get install build-essential autoconf automake libtool pkg-config
$ sudo apt-get install nmap trace-cmd sshpass kernelshark net-tools
$ sudo apt-get install python-matplotlib python-numpy libfreetype6-dev libpng12-dev python-nose
$ sudo apt-get install python-pip python-dev
$ sudo pip install --upgrade trappy bart-py devlib
NOTE: TRAPpy and BART depend on ipython and ipython-notebook. Some IPython Notebooks examples are written in JSON nbformat version 4 which might not be supported by the IPython version installed by apt-get (current version is 1.2.1-2 which does not support such format). In this case, it is needed to remove IPython and install it using pip instead:
$ sudo apt-get remove ipython ipython-notebook
$ sudo pip install ipython ipython-notebook
The code of the LISA toolkit with all the supported tests and Notebooks can be cloned from the official GitHub repository with this command:
$ git clone https://github.com/ARM-software/lisa.git
The target platform to be used for experiments with LISA must satisfy the following requirements:
- allow ssh access, preferably as root, using either a password or an SSH key
- support sudo, even if it's accessed as root user
- allow adb access, eventually by specifying a DEVICE ID
- the local shell should define the ANDROID_HOME environment variable pointing to an Android SDK installation
Most of the tests targets a kernel with support for some new frameworks which are currently in-development:
- Energy-Aware Scheduler (EAS)
- SchedFreq: the CPUFreq governor
- SchedTune: the central, scheduler-driven, power-perfomance control
Tests targeting an evaluation of these frameworks requires also a set of tracepoint which are not available in mainline kernel. The series of patches required to add to a recent kernel the tracepoints required by some tests are available on this git repository:
git://www.linux-arm.org/linux-power.git lisa/debug
The patches required are: lisa/debug_base..lisa/debug
An easy way to test your installation is to give a run to the EAS RFC tests. These are a set of experiments which allows to compare EAS performance and energy consumption with respect to standard kernel.
NOTE: The following tutorial is still recommended it you want to get a better grasp on how the framework is organized and how to use it at your best.
Let's assume your target is running an EAS enabled kernel, to run such tests just run these few steps:
-
Check the target.config file in the root folder of the toolkit
In that file you to specify the proper value for at least platform and board, as well as the host IP address of the target and the login credentials (i.e. username and password). -
Setup the execution environment by sourcing the provided initialization file from the shell you will use to run the experiments
$ source init_env -
Run the EAS RFC test using the standard nosetest command:
$ nosetests -v tests/eas/rfc.py
-
Wait for the test to complete and than you can report the results with:
./tools/report.py --base noeas --tests eas
The toolkit provides these resources:
A collection of "assets" required to run examples and tests:
.
|-- assets
| `-- mp3-short.json
A script to setup the test environment.
|-- init_env
A collection of IPython Notebooks grouped by topic:
|-- ipynb
| |-- devlib
| |-- ipyserver_start
| |-- ipyserver_stop
| |-- nohup.out
| |-- sched_dvfs
| |-- utils
| `-- wlgen
A set of support libraries to easily develop new IPython notebooks and tests:
|-- libs
| |-- __init__.py
| |-- utils
| `-- wlgen
A JSON based configuration for the target device to use for running the tests:
|-- target.config
A collection of regression tests grouped by topic:
|-- tests
| `-- eas
A collection of binary tools, pre-compiled for different architectures, which are required to run tests on the target device:
`-- tools
|-- arm64
|-- armeabi
|-- plots.py
|-- report.py
|-- scripts
`-- x86
(these are provided in binary form because they correspond to specific known-good versions and it avoids having to cross-compile these tools for the target)
This section provides a quick start guide on understanding the toolkit by guiding the user though a set of example usage scenarios.
IPython is a web based interactive python programming interface. This toolkit provides a set of IPython notebooks ready to use. To use these notebooks, an IPython server must be started to serve pages to a browser on the same host machine or over a local network.
# Enter the ipynb folder
$ cd ipynb
# Start the server
$ ./ipyserver_start lo
This will start the IPython server and open the index page in a new browser tab. If the index is not automatically loaded in the browser, visit the link reported by the server startup script.
The index page is an HTML representation of the local ipynb folder. From that page we can access all the IPython notebooks provided by the toolkit.
Typical notebooks and tests will make use of the TestEnv class to initialize and access a remote target device. An overall view of the functionality exposed by this class can be seen in this notebook: utils/testenv_example.ipynb
RT-App is a configurable synthetic workload generator used to run different intensity experiments on a target. LISA provides a python wrapper API to simplify the definition of RT-App based workloads and their execution on a target.
This notebook: wlgen/simple_rtapp.ipynb is a complete example of setting up an experiment, execution and data collection.
Specifically it demonstrates how to:
- configure a target for an experiment
- configure FTrace for event collection
- configure an HWMon based energy meter for energy measurements
- configure a simple rt-app based workload consisting of two different tasks
- run the workload on the target to collect FTrace events and energy consumption
- visualize scheduling events using the in-browser trace plotter provided by the TRAPpy library
- visualize some simple performance metrics for the tasks
Tests are usually performed on a set of data collected while executing properly defined experiments. An experiment is usually defined by a specific target configuration which is used to run a certain workload mix. Thus, to run experiments we usually need the support of a proper module which configure a target and execute a workload on it in order to collect the data required for a test.
The Executor module is a simple yet effective support to collect all the data required for a test. This notebook: utils/executor_example.ipynb is a simple example of how to use the Executor collect experimental data for a predefined set of target configurations and by running a specified set of workloads.
One of the main aims of LISA is to become a repository for regression tests on scheduler and power management behavior. A common pattern for defining new test cases is to start with an IPython notebook to design an experiment and compute metrics of interest. The notebook can then be converted into a self-contained test to run in batch mode.
An example of such a notebook is: sched_dvfs/smoke_test.ipynb
sched-DVFS is a technique for scheduler driven DVFS operation and is a key part of Energy Aware Scheduling (EAS). LISA is used extensively for EAS analysis and some of the examples listed in this README are taken from the EAS testing and evaluation experience. To know more about EAS and sched-DVFS, see: (http://www.linaro.org/blog/core-dump/energy-aware-scheduling-eas-progress-update/)
In this notebook the toolkit API is more extensively used to define an experiment to:
- select and configure three different CPUFreq governors
- run a couple of RTApp based test workloads in each configuration
- collect and plot scheduler and CPUFreq events
- collect and compare the energy consumption during workload execution in each of the different configurations
The notebook compares three different CPUFreq governors: "performance", "sched" and "ondemand". New configurations are easy to add. For each configuration the notebook generates plots and tabular reports regarding working frequencies and energy consumption.
This notebook is a good example of using LISA to build a new set of experiments which can then be transformed into a standalone regression test.
Once a new set of tests have been defined and verified, perhaps by using a notebook to develop them, they can be transformed into a standalone regression test. Regression tests are written using the same API used to develop a notebook, thus their transformation into a batch task is generally quite easy, especially considering that a notebook can be exported as a standalone python script.
An example of such a regression test is:
tests/eas/rfc.py
This test, which is used for EAS analysis, is designed to allow different configurations and workloads to be compared from both a performance and an energy standpoint.
To run this regression test, first set up the local execution environment by sourcing the initialization script:
$ source init_env
Next, check the target configuration which is defined in target.config. This file has to be updated to at least define the login credentials for the target to use, and the kind of platform (e.g. "linux" or "android") and the board (if it is supported by the toolkit, e.g. "tc2" or "juno")
The set of target configurations considered by the test as well as the set of workloads to execute with each configuration is defined by a test specific configuration file. In the case of the EAS regression suite, this file is tests/eas/rfc_eas.config. Have a look at this file and ensure to enable/tune the "confs" and "wloads" sections. The default configuration runs a predefined set of tests which are commonly used for EAS RFC postings.
Once eveything has been configured to run the test, execute it with:
nosetests -v tests/eas/rfc.py:EAS
This command will run all the configured experiments and collect the results into the output folder generated by the TestEnv and pointed to by the "results_latest" symlink in the top folder.
Once the test has completed, report the results using the command:
./tools/report.py --base noeas --tests eas
This will generate a table comparing energy/performance metrics for the "eas" configuration with respect to the "noeas" configuration.
Regression tests make use of the test environment generated by the TestEnv module. By default, this module configures the target defined by the target.conf file present at the top level folder.
The comments in this file should be good enough to understand how to properly setup a target to be used for the execution of all the tests provided by the toolkit.
The configuration of a specific regression test is usually provided by a corresponding configuration file. For example, the configuration file for the tests/eas/rfc.py tests is provided by the tests/eas/rfc.conf.
This configuration file describes:
- which devlib modules are required by this experiment
- which binary tools need to be deployed in the target to run the experiments
- other devlib specific configurations (e.g. FTrace events of interest)
- the set of target configurations (confs) to test
- the set of workloads (wloads) to run
The test will run each workload with each specified target kernel configuration.
The results of a specific experiment execution can be obtained once the test has completed using this command:
./tools/report.py --bases <regexp1> --tests <regexp2>
This script compares a base configuration (whose name matches the regular expression regexp1), to each target configuration which name matches the regexp2 regular expression.
A default set of plots can be generated for all the executed experiments using this command:
./tools/plot.py
This script will produce, for each run of the experiments, a set of plots saved as PNG images into the output folder of the experiment.
This project is licensed under Apache-2.0.
This project includes some third-party code under other open source licenses. For more information, see lisa/tools/LICENSE.*
Contributions are accepted under Apache-2.0. Only submit contributions where you have authored all of the code. If you do this on work time make sure your employer is cool with this.