On 03/29/2016 04:11 AM, Lukáš Doktor wrote:
Dne 28.3.2016 v 21:49 Cleber Rosa napsal(a):
----- Original Message -----
From: "Cleber Rosa" <cr...@redhat.com>
To: "Lukáš Doktor" <ldok...@redhat.com>
Cc: "Amador Pahim" <apa...@redhat.com>, "avocado-devel"
<avocado-devel@redhat.com>, "Ademar Reis" <ar...@redhat.com>
Sent: Monday, March 28, 2016 4:44:15 PM
Subject: Re: [Avocado-devel] RFC: Multi-host tests
----- Original Message -----
From: "Lukáš Doktor" <ldok...@redhat.com>
To: "Ademar Reis" <ar...@redhat.com>, "Cleber Rosa"
<cr...@redhat.com>,
"Amador Pahim" <apa...@redhat.com>, "Lucas
Meneghel Rodrigues" <look...@gmail.com>, "avocado-devel"
<avocado-devel@redhat.com>
Sent: Saturday, March 26, 2016 4:01:15 PM
Subject: RFC: Multi-host tests
Hello guys,
Let's open a discussion regarding the multi-host tests for avocado.
The problem
===========
A user wants to run netperf on 2 machines. To do it manually he does:
machine1: netserver -D
machine1: # Wait till netserver is initialized
machine2: netperf -H $machine1 -l 60
machine2: # Wait till it finishes and report store the results
machine1: # stop the netserver and report possible failures
Now how to support this in avocado, ideally as custom tests, ideally
even with broken connections/reboots?
Super tests
===========
We don't need to do anything and leave everything on the user. He is
free to write code like:
...
machine1 = aexpect.ShellSession("ssh $machine1")
machine2 = aexpect.ShellSession("ssh $machine2")
machine1.sendline("netserver -D")
# wait till the netserver starts
machine1.read_until_any_line_matches(["Starting netserver"], 60)
output = machine2.cmd_output("netperf -H $machine1 -l
$duration")
# interrupt the netserver
machine1.sendline("\03")
# verify netserver finished
machine1.cmd("true")
...
the problem is it requires active connection and the user needs to
manually handle the results.
And of course the biggest problem here is that it doesn't solve the
Avocado problem: providing a framework and tools for tests that span
multiple (Avocado) execution threads, possibly on multiple hosts.
Well it does, each "ShellSession" is a new parallel process. The only
problem I have with this design is that it does not allow easy code
reuse and the results strictly depend on the test writer.
Yes, *aexpect* allows parallel execution in an asynchronous fashion. Not
targeted to tests *at all*. Avocado, as a test framework, should deliver
more. Repeating the previous wording, it should be "providing a
framework and tools for tests that span multiple (Avocado) execution
threads, possibly on multiple hosts."
Triggered simple tests
======================
Alternatively we can say each machine/worker is nothing but yet
another
test, which occasionally needs a synchronization or data-exchange. The
same example would look like this:
machine1.py:
process.run("netserver")
barrier("server-started", 2)
barrier("test-finished", 2)
process.run("killall netserver")
machine2.py:
barrier("server-started", 2)
self.log.debug(process.run("netperf -H %s -l 60"
% params.get("server_ip"))
barrier("test-finished", 2)
where "barrier(name, no_clients)" is a framework function which makes
the process wait till the specified number of processes are waiting
for
the same barrier.
The barrier mechanism looks like an appropriate and useful utility
for the
example given. Even though your use case example explicitly
requires it,
it's worth pointing out and keeping in mind that there may be valid
use cases
which won't require any kind of synchronization. This may even be
true to
the executions of tests that spawn multiple *local* "Avocado runs".
Absolutely, this would actually allow Julio to run his "Parallel
(clustered) testing".
So, let's try to identify what we're really looking for. For both the
use case I mentioned and Julio's "Parallel (clustered) testing", we need
a (the same) test run by multiple *runners*. A runner in this context is
something that implements the `TestRunner` interface, such as the
`RemoteTestRunner`:
https://github.com/avocado-framework/avocado/blob/master/avocado/core/remote/runner.py#L37
The following (pseudo) Avocado Test could be written:
from avocado import Test
# These are currently private APIs that could/should be or
# be exposed under another level. Also, the current API is
# very different from what is used here, please take it as
# pseudo code that might look like a future implementation
from avocado.core.remote.runner import RemoteTestRunner
from avocado.core.runner import run_multi
from avocado.core.resolver import TestResolver
from avocado.utils.wait import wait_for
class Multi(Test):
def test(self):
worker1 = RemoteTestRunner('worker1')
worker2 = RemoteTestRunner('worker2')
# Resolve a local test to send it to be run on multiple machines
test = TestResolver().resolve('bonnie.py')
# run_multi is asynchronous, and results can be queried about its status
results = run_multi([worker1, worker2], test)
wait_for(results.finished, self.timeout)
# combine remote whiteboard (with performance results) keyed by worker
name
whiteboard = {}
for worker_result in results:
whiteboard[worker_result.name] = worker_result.whiteboard
self.whiteboard = whiteboard
If any kind of synchronization was necessary between workers, the
barrier utility library could be used, maybe even
transparently as part of "run_multi". Parameter passing to tests is also
a layered issue. My point is that this seems to
include the features needed to allow "tests that span multiple machines".
Does it look reasonable?
A lot of sugar coating can (and should) be added on top. Creating
workers automatically, having a superclass for tests
that span multiple machines, plugins that take worker names directly
from command line options and what not are
likely natural additions.
The barrier needs to know which server to use for communication so we
can either create a new service, or simply use one of the
executions as
"server" and make both processes use it for data exchange. So to
run the
above tests the user would have to execute 2 avocado commands:
avocado run machine1.py --sync-server machine1:6547
avocado run machine2.py --remote-hostname machine2 --mux-inject
server_ip:machine1 --sync machine1:6547
where:
--sync-server tells avocado to listen on ip address machine1
port 6547
--remote-hostname tells the avocado to run remotely on machine2
--mux-inject adds the "server_ip" into params
--sync tells the second avocado to connect to machine1:6547 for
synchronization
To be honest, apart from the barrier utility, this provides little
value
from the PoV of a *test framework*, and possibly unintentionally,
competes
and overlaps with "remote" tools such as fabric.
Also, given that the multiplexer is an optional Avocado feature, such
a feature should not depend on it.
It does not, these are only used to demonstrate this particular
feature. You can hardcode the values in the tests, you can use
env-variables or any other feature.
Basically this "mht" format is nothing more, than list of "avocado
run" commands to be executed in parallel and it's focus was on
simplicity, maybe even only for demonstration purposes.
Running those two tests has only one benefit compare to the previous
solution and that is it gathers the results independently and makes
allows one to re-use simple tests. For example you can create a 3rd
test, which uses different params for netperf, run it on "machine2"
and
keep the same script for "machine1". Or running 2 netperf senders
at the
same time. This would require libraries and more custom code when
using
"Super test" approach.
There are additional benefits for this solution. When we introduce the
locking API, tests running on a remote machine will be actually
directly
executed in avocado, therefor the locking API will work for them,
avoiding problems with multiple tests using the same shared resource.
Another future benefit would be system reboot/lost connection when we
introduce this support for individual tests. The way it'd work is that
user triggers the jobs, the master remembers the test ids and would
poll
for results until they finish/timeout.
All of this we get for free thanks to re-using the existing
infrastructure (or the future infrastructure), so I believe this is
the
right way to go and in this RFC I'm describing details of this
approach.
All of the benefits listed are directly based on the fact that tests on
remote systems would be run under the Avocado test runner and would
have
it's runtime libraries available. This is a valid point, but again it
doesn't bring a significant change in the user experience wrt running
tests that span multiple "Avocado runs" (possibly on remote machines).
Basically this is the key part of this RFC. I like the idea of running
avocado processes for each test, instead of yet another remote
execution handling. The biggest benefit are the test results in well
known format and the possibility to run/combine all the tests
supported by avocado.
Actually I have avocado-in-avocado script in my CI testing, it just
waits for the long-names fix to be applied as it generates too long
test names. But I tested it with the fix and the results are very nice
and easy to analyze as you simply go through results you know from
simple testing.
Triggering the jobs
-------------------
Previous example required the user to run the avocado 2 times (per
each
machine) and sharing the same sync server. Additionally it resulted
into
2 separated results. Let's try to eliminate this problem.
Basic tests
~~~~~~~~~~~
For basic setups, we can come up with very simple format to describe
which tests should be triggered and avocado should take care of
executing it. The way I have in my mind is to simply accept list of
"avocado run" commands:
simple_multi_host.mht:
machine1.py
machine2.py --remote-hostname machine2 --mux-inject
server_ip:machine1
Running this test:
avocado run simple_multi_host.mht --sync-server 0.0.0.0
avocado would pick a free port and start the sync server on it.
Then it
would prepend "avocado run" and append "--sync $sync-server
--job-results-dir $this-job-results" to each line in
"simple_multi_host.mht" and run them in parallel. Afterward it'd wait
till both processes finish and report pass/fail depending on the
status.
This way users get overall results as well as individual ones and
simple
way to define static setups.
First, the given usage example would require Avocado to introduce:
* A brand new file format
* A new test type (say MULTI_HOST_TEST, in addition to the SIMPLE,
INSTRUMENTED, etc).
Introducing a brand new file format may look like a very simple thing
to do, but it's not. I can predict that we'd learn very quickly that
our original file format definition is very limited. Then we'd either
have to live with that, or introduce new file format versions, or just
break the initial definition or compatibility. These are all problems
related to file formats, not really to your proposed file format.
Then, analogous to the "remote tools (fabric)" example I gave before,
this looks to be outside of the problem scope of Avocado, in the sense
that "template" tools can do it better.
Introducing a new test type, and a test resolver/loader, would be a
mandatory step to achieve this design, but it looks like a necessary
action only to make the use of "MHT" file format possible.
Please note that having a design that allow users to fire multiple
Avocado command line instances executions in their own scripts is a bad
thing, but as a test framework, I believe we can deliver a better, more
focused experience.
I meant "is *not* a bad thing".
I think you have a point here. My idea was to support new-line
separated list of avocado executions as a simple wrapper to run
processes in parallel as it's very simple to develop and it's not
promising anything. It simply takes whatever you hand it over, spawns
multiple processes and gives you results.
Then to add some value I added the --sync handling as it's one
problematic thing. Basically it can be written in a generic way, but I
see your point with hard-to-debug failures or unexpected behavior.
It was meant to be a very simple and easy to understand way to promote
multi-host-testing but it can as well become very painful thing if
people start relying on it. So maybe we should only introduce the real
thing below.
Contrib scripts
~~~~~~~~~~~~~~~
The beauty of executing simple lines is, that users might create
contrib
scripts to generate the "mht" files to get even better flexibility.
Since I don't think a new file format and test type is a good thing,
this
also becomes a bad idea IMHO.
Advanced tests
~~~~~~~~~~~~~~
The above might still not be flexible enough. But the system
underneath
is very simple and flexible. So how about creating instrumented tests,
which generate the setup? The same simple example as before:
multi_host.py
runners = ["machine1.py"]
runners.append("machine2.py --remote-hostname machine2
--mux-inject
server_ip:machine1")
self.execute(runners)
A major plus here is that there's no attempt to define new file
formats,
test types and other items that are necessary only to fulfill a use
case
requirement. Since Avocado's primary language of choice is Python, we
should stick to it, given that it's expressive enough and well
maintained
enough. This is of course a lesson we learned with Autotest itself,
let's
not forget it.
Then, a couple of things I dislike here:
1) First runner is special/magical (sync server would be run here)
2) Interface with runner execution is done by command line parameters
Well the 0-st runner is special (the one which executes the
multi-host-instrumented-test). It needs to listen on any free port and
pass this port to all executed tests (if they use barriers/sync).
I'll talk about the 2nd point later....
where the "self.execute(tests)" would take the list and does the
same as
for basic tests. Optionally it could return the json results per each
tests so the test itself can react and modify the results.
The above was just a direct translation of the previous example,
but to
demonstrate the real power of this let's try a PingPong multi host
test:
class PingPong(MultiHostTest):
def test(self):
hosts = self.params.get("hosts", default="").split(";")
assert len(hosts) >= 2
runners = ["ping_pong --remote-hostname %s" % _
for _ in hosts]
# Start creating multiplex tree interactively
mux = MuxVariants("variants")
# add /run/variants/ping with {} values
mux.add("ping", {"url": hosts[1], "direction": "ping",
"barrier": "ping1"})
# add /run/variants/pong with {} values
mux.add("pong", {"url": hosts[-1], "direction": "pong",
"barrier": "ping%s" % len(hosts) + 1})
# Append "--mux-inject mux-tree..." to the first command
runners[0] += "--mux-inject %s" % mux.dump()
for i in xrange(1, len(hosts)):
mux = MuxVariants("variants")
next_host = hosts[i+1 % len(hosts)]
prev_host = hosts[i-1]
mux.add("pong", {"url": prev_host, "direction":
"pong",
"barrier": "ping%s" % i})
mux.add("ping", {"url": next_host, "direction":
"ping",
"barrier": "ping%s" % i+1})
runners[i] += "--mux-inject %s" % mux.dump()
# Now do the same magic as in basic multihost test on
# the dynamically created scenario
self.execute(runners)
The `self.execute` generates the "simple test"-like list of "avocado
run" commands to be executed. But the test writer can define some
additional behavior. In this example it generates
machine1->machine2->...->machine1 chain of ping-pong tests.
You mean that this would basically generate a "shell script like" list
of avocado runs? This looks to be a very strong design decision, and
I fail to see how it would lend itself to be flexible enough and
deliver
the "test writer can define some additional behavior" requirement.
Explanation below...
When running "avocado run pingpong --mux-inject
hosts:machine1;machine2"
this generates 2 jobs, both running just a single "ping_pong" test
with
2 multiplex variants:
machine1:
variants: !mux
ping:
url: machine2
direction: pong
barrier: ping1
pong:
url: machine2
direction: pong
barrier: ping2
machine2:
variants: !mux
pong:
url: machine1
direction: pong
barrier: ping1
ping:
url: machine1
direction: ping
barrier: ping2
The first multiplex tree for three machines looks like this:
variants: !mux
ping:
url: machine2
direction: pong
barrier: ping1
pong:
url: machine3
direction: pong
barrier: ping
Btw I simplified the format for the sake of this RFC. I think
instead of
generating the strings we should support API to specify test,
multiplexer, options... and then turn them into the parallel executed
jobs (usually remotely). But these are just details to be solved if we
decide to work on it.
This statement completely changes what you have proposed up to this
point.
IMHO it's far from being just details, because that would define the
lowest
and commonest level of this feature set that we would advertise and
support.
The design should really be from this level up, and not from the
opposite
direction.
If external users want to define file formats (say your own MHT
proposal) on
top of our "framework for running tests that span multiple execution
threads"
at once, they should be able to do so.
If you ask me, having sound Avocado APIs that users could use to fire
multiple
portions of their *tests* at once and have their *results* coalesced
into a
single
*test* result is about what Avocado should focus on.
And this was suppose to be the answer. In the end yes, I think it
should generate the "avocado run" command with result-dir based inside
this test's results. The reason is it gives you the results you know
per each worker and they can run independently (survive the network
issues, system reboots when we add the support for it in avocado)
This would really be an implementation detail of the chosen runner. The
current remote runner actually run Avocado, but that's its
(RemoteTestRunner) own internal design decision. It does have a lot of
pluses, but that is not the focus of this conversation. Another runner,
say, ThinTestRunner, could choose to do things differently.
Having said that, I completely agree that we should, unless proven
wrong, reuse the RemoteTestRunner for multi-host tests.
The alternative is to create a client worker, which executes code on
demand, but that's more complex and it'd double the effort if we
decide to support system reboots/connection issues.
Agreed. Having an agent/broker on the remote side does not seem to be
necessary or beneficial at this point.
What this paragraph was about is that it should not probably directly
generate the arguments, but we should define an API which adds
individual pieces of information and is translated into the command at
the end.
I decided not to go into details here as I thought it's better to
focus on part1 (--sync --sync-server) which already has a proof of
concept version out there. Then I wanted to create the "mht" file,
which would demonstrate how the results could look like, and how it
all goes together and when we have those results and issues, we can
introduce the instrumented-test API which would evolve from the
real-world issues.
Results and the UI
==================
The idea is, that the user is free to run the jobs separately, or to
define the setup in a "wrapper" job. The benefit of using the
"wrapper"
job are the results in one place and the `--sync` handling.
The difference is that running them individually looks like this:
1 | avocado run ping_pong --mux-inject url:192.168.1.58:6001
--sync-server
1 | JOB ID : 6057f4ea2c99c43670fd7d362eaab6801fa06a77
1 | JOB LOG :
/home/medic/avocado/job-results/job-2016-01-22T05.33-6057f4e/job.log
1 | SYNC : 0.0.0.0:6001
1 | TESTS : 1
1 | (1/1) ping_pong: \
2 | avocado run ping_pong --mux-inject :url::6001 direction:pong
--sync 192.168.1.1:6001 --remote-host 192.168.1.1
2 | JOB ID : 6057f4ea2c99c43670fd7d362eaab6801fa06a77
2 | JOB LOG :
/home/medic/avocado/job-results/job-2016-01-22T05.33-6057f4e/job.log
2 | TESTS : 1
2 | (1/1) ping_pong: PASS
1 | (1/1) ping_pong: PASS
and you have 2 results directories and 2 statuses. By running them
wrapped inside simple.mht test you get:
avocado run simple.mht --sync-server 192.168.122.1
JOB ID : 6057f4ea2c99c43670fd7d362eaab6801fa06a77
JOB LOG :
/home/medic/avocado/job-results/job-2016-01-22T05.33-6057f4e/job.log
TESTS : 1
(1/1) simple.mht: PASS
RESULTS : PASS 1 | ERROR 0 | FAIL 0 | SKIP 0 | WARN 0 |
INTERRUPT 0
TIME : 0.00 s
And single results:
$ tree $RESULTDIR
└── test-results
└── simple.mht
├── job.log
...
├── 1
│ └── job.log
...
└── 2
└── job.log
...
tail -f job.log:
running avocado run ping pong ping pong
running avocado run pong ping pong ping --remote-hostname
192.168.122.53
waiting for processes to finish...
PASS avocado run ping pong ping pong
FAIL avocado run pong ping pong ping --remote-hostname
192.168.122.53
this job FAILED
I won't spend much time here, since the UI is bound to follow other
design
ideas/decisions.
Sure, the important part here is the results format.
Demonstration
=============
While considering the design I developed a WIP example. You can
find it
here:
https://github.com/avocado-framework/avocado/pull/1019
It demonstrates the `Triggered simple tests` chapter without the
wrapping tests. Hopefully it helps you understand what I had in
mind. It
contains modified "examples/tests/passtest.py" which requires 2
concurrent executions (for example if you want to test your server and
run multiple concurrent "wget" connections). Feel free to play with
it,
change the number of connections, set different barriers, combine
multiple different tests...
Autotest
========
Avocado was developed by people familiar with Autotest, so let's just
mention here, that this method is not all that different from Autotest
one. The way Autotest supports parallel execution is it let's users to
create the "control" files inside the multi-host-control-file and then
run those in parallel. For synchronization it contains master->slave
barrier mechanism extended of SyncData to send pickled data to all
registered runners.
I considered if we should re-use the code, but:
1. we do not support control files, so I just inspired by passing the
params to the remote instances
One of the wonderful things about Autotest control files is that
it's not a custom file format. This can not be underestimated. While
other frameworks have had huge XML based file formats to drive their
jobs, Autotest control files are infinitely more capable and their
readability is a lot more scalable.
The separation of client and server test types (and control files) is
actually what prevents control files from nearing perfection IMHO.
Yep
The server API allows you to run client control files on given hosts.
These client control files usually need tweaking for each host. Then
you're suddenly doing code generation (control files Python code). That
is not nice.
The tests I saw usually generated simple "runTest" with different
params. So what I'm proposing is actually similar, let's run avocado
and allow params passing.
I believe that, if Avocado provides such an API that allows regular
Python
code to operate similarly to server control files, while giving more
control
and granularity to what is run on the individual job executions (say
on remote machines), and help to coalesce the individual portions
into a
single test result, it would be a very attractive tool.
I think the multi-host test should only pick existing normal tests and
run the set of tests they need to perform the task using barriers to
synchronize it.
Actually there is one thing which is significantly limiting the usage
and that's the multiplexer. I'd like to run:
"avocado run boot migrate recievemigrate migrate recievemigrate
shutdown" tests and use different params for each tests. Currently
this is not possible and it's something I'd been proposing all the
time. (mapping params to individual tests).
Anyway even without this mapping we can do all kinds of setups and
when we add such feature we can always start using it in
multi-host-testing as multi-host-testing is just triggering
avocado-jobs in terms of this RFC so all features available in avocado
are available to each worker in multi-host-testing.
PS: The multiplexer is not needed for multi-host-tests, you're free to
hard-code the values inside tests or to use whatever way to tell the
test what it should do. The barriers are using the server from
"--sync" cmdline argument so the test is the only component which
might need to be parametric.
I will, on purpose, not explore the parameter passing problems until we
are more or less on the same page about the bare bones of "running a
test that span multiple machines". Then we can explore this optional but
very important aspect.
2. the barriers and syncdata are quite hackish, master->slave
communication. I think the described (and demonstrated) approach does
the same in a less hackish way and is easy to extend
Using this RFC we'd be able to run autotest-multi-host tests, but it'd
require rewriting the control files to "mht" (or contrib) files.
It'd be
probably even possible to write a contrib script to run the control
file
and generate the "mht" file which would run the autotest test. Anyway
the good think for us is, that this does not affect "avocado-vt",
because all of the "avocado-vt" multi-host tests are using a single
"control" file, which only prepares the params for simple avocado-vt
executions. The only necessary thing is a custom "tests.cfg" as by
default it disallows multi-host tests (or we can modify the
"tests.cfg"
and include the filter inside the "avocado-vt" loader, but these are
just the details to be sorted when we start running avocado-vt
multi-host tests.
Conclusion
==========
Multi-host testing was solved many times in the history. Some hardcode
tests with communication, but most framework I had seen support
triggering "normal/ordinary" tests and add some kind of barrier
(either
inside the code or between the tests) mechanism to synchronize the
execution. I'm for the flexibility and easy test sharing and that
is how
I described it here.
Kind regards,
Lukáš
_______________________________________________
Avocado-devel mailing list
Avocado-devel@redhat.com
https://www.redhat.com/mailman/listinfo/avocado-devel
_______________________________________________
Avocado-devel mailing list
Avocado-devel@redhat.com
https://www.redhat.com/mailman/listinfo/avocado-devel
