-
Notifications
You must be signed in to change notification settings - Fork 0
ControlHowto
The key defining component of a job is its control file; this file defines all aspects of a jobs life cycle. The control file is a python script which directly drives execution of the tests in question.
You are automatically supplied with a job object which drives the job and supplies services to the control file. A control file can be as simple as:
job.run_test('kernbench')
The only mandatory argument is the name of the test. There are lots of examples; each test has a sample control file under tests/<testname>/control
If you're sitting in the top level of the autotest client, you can run the control file like this:
$ bin/autotest <control_file_name>
($ is just the shell prompt - you don't need that bit ;-)) You can also supply specific arguments to the test
job.run_test('kernbench', iterations=2, threads=5)
- First paramater is the test name
- The others are arguments to the test. Most tests will run with no arguments if you want the defaults
If you would like to specify a tag for the results directory for a particular test:
job.run_test('kernbench', iterations=2, threads=5, tag='mine')
Will create a results directory "kernbench.mine" instead of the default "kernbench". This is particularly important when writing more complex control files that may run the same test multiple times, in order to properly separate the results of each of the test runs they will need a unique tag.
Sometimes when you are developing a test it's useful to have it packaged somewhere so your control file can download it, uncompress it and run it. The convention for packaging test is on ExternalTests. Make sure you read that session before you try to package and run your own external tests.
One of the benefits of the use of a true programming language for the control script is the ability to use all of its structural and error checking features. Here we use a loop to run kernbench with different threading levels.
for t in [8, 16, 32]:
job.run_test('kernbench', iterations=2, threads=t, tag='%d' % t)
After every reboot and after every test, Autotest will collect a variety of standard pieces of system information made up of specific files grabbed from the filesystem (e.g. /proc/meminfo) and the output of various commands (e.g. "uname -a"). You can see this output in the results directories, under sysinfo/ (for per-reboot data) and <testname>/sysinfo (for pre-test data).
For a full list of what's collected by default you can take a look at client/bin/base_sysinfo.py; however, there also exists a mechanism for adding extra files and commands to the system info collection inside your control files. To add a custom file to the log collection you can call:
job.add_sysinfo_file("/proc/vmstat")
This would collect the contents of /proc/vmstat after every reboot. To collect it on every test you can use the optional on_every_test parameter, like so:
job.add_sysinfo_file("/proc/vmstat", on_every_test=True)
There exists a similar method for adding a new command to the sysinfo collection:
job.add_sysinfo_command("lspci -v", logfile="lspci.txt")
This will run "lspci -v" through the shell on every reboot, logging the output in lspci.txt. The logfile parameter is optional; if you do not specify it, the logfile will default to the command text with all whitespace replaced with underscores (e.g. in this case it would use lspci_-v as the filename). This method also takes an on_every_test parameter that can be used to run the collection after every test instead of every reboot.
You can enable one or more profilers to run during the test. Simply add them before the tests, and remove them afterwards, eg:
job.profilers.add('oprofile')
job.run_test('sleeptest')
job.profilers.delete('oprofile')
If you run multiple tests like this:
job.profilers.add('oprofile')
job.run_test('kernbench')
job.run_test('dbench')
job.profilers.delete('oprofile')
It will create separate profiling output for each test - generally we do a separate profiling run inside each test, so as not to perturb the performance results. Profiling output will appear under <testname>/profiling in the results directory.
Again, there are examples for all profilers in profilers/<profiler-name>/control.
We have support built in for creating filesystems. Suppose you wanted to run the fsx test against a few different filesystems:
# Uncomment this line, and replace the device with something sensible
# for you ...
# fs = job.filesystem('/dev/hda2', job.tmpdir)
for fstype in ('ext2', 'ext3'):
fs.mkfs(fstype)
fs.mount()
try:
job.run_test('fsx', job.tmpdir, tag=fstype)
finally:
fs.unmount()
or if we want to show off and get really fancy, we could mount ext3 with a bunch of different options, and see how the performance compares across them:
fs = job.filesystem('/dev/sda3', job.tmpdir)
iters=10
for fstype, mountopts, tag in (('ext2', '', 'ext2'),
('ext3', '-o data=writeback', 'ext3writeback'),
('ext3', '-o data=ordered', 'ext3ordered'),
('ext3', '-o data=journal', 'ext3journal')):
fs.mkfs(fstype)
fs.mount(args=mountopts)
try:
job.run_test('fsx', job.tmpdir, tag=tag)
job.run_test('iozone', job.tmpdir, iterations=iters, tag=tag)
job.run_test('dbench', iterations=iters, dir=job.tmpdir, tag=tag)
job.run_test('tiobench', dir=job.tmpdir, tag=tag)
finally:
fs.unmount()
Where a job needs to cause a system reboot such as when booting a newly built kernel, there is necessarily an interuption to the control script execution. The job harness therefore also provides a phased or step based interaction model.
def step_init():
job.next_step([step_test])
testkernel = job.kernel('2.6.18')
testkernel.config('http://mbligh.org/config/opteron2')
testkernel.build()
testkernel.boot() # does autotest by default
def step_test():
job.run_test('kernbench', iterations=2, threads=5)
job.run_test('dbench', iterations=5)
By defining a step_init this control script has indicated it is
using step mode. This triggers automatic management of the step state
across breaks in execution (such as a reboot) maintaining forward flow.
In broad brush, the control file is not simply executed once, but repeatedly executed until it indicates the job is complete. In a stand-alone context we would expect to re-start execution automatically on boot when a control file exists, in a managed environment the managing server would perform the same role.
Obviously looping is more difficult in the face of phase based execution. The state maintained by the stepping engine is such, that we can implement a boot based loop using step parameters.
def step_init():
step_test(1)
def step_test(iteration):
if (iteration < 5):
job.next_step([step_test, iteration + 1])
print "boot: %d" % iteration
job.run_test('kernbench', tag="%d" % i)
job.reboot()
The job object also provides a parallel method for running multiple tasks at the same time. The method takes a variable number of arguments, each representing a different task to be run in parallel. Each argument should be a list, where the first item on the list is a function to be called and all the remaining elements are arguments that will be passed to the function when it is called.
def first_task():
job.run_test('kernbench')
def second_task():
job.run_test('dbench')
job.parallel([first_task], [second_task])
This control file will run both kernbench and dbench at the same time. Alternatively, this could've been written as:
job.parallel([job.run_test, 'kernbench'], [job.run_test, 'dbench'])
However, if you want to so something more complex in your tasks than call a single function then you'll have to define your own functions to do it, as in the first example.
The parallel jobs are run through fork, so each task will be running in its own address space and you don't need to worry about performing any process-local synchronization between your separate tasks. However, these processes will still be running on the same machine and so still need to make certain that these tasks don't crash into each other while accessing shared resources (e.g. the filesystem). This means no rebooting during parallel tasks, and if you're running the same test in different tasks, you must be sure to give each task a unique tag