I’ve done some research on this and I’m fairly convinced it doesn’t
exist,
but I thought I would just ask here in case I missed something.
I want to be able to full-stack load test a Ruby on Rails site (not that
rails matters in this context) running dynamic test plans across
multiple
concurrent users in order to better replicate real world usage.
In other words, I’d like to run say 50 concurrent user sessions where
some
of them are signing up, some are just browsing around, some are logging
in
and doing UserActivity1, some are doing UserActivity2, etc.
I want to script the test plans in ruby, similar to a seleinum-rc test
plan
where I can use the same DSL to simulate user input and identify parts
of
the response html, and use that to drive further test activity. I’d like
to
use some of the same DSL from my selenium or integration test helper
modules
already developed in these tests as well in order to be DRY.
I would like the test framework to mimc the browser behaviour: handling
cookies and even simulating caching to send the appropriate
If-Modified-Since headers.
Since I would like to replicate actual request load on the server, I
want to
make sure that all page assets, even static ones like
images/css/js/flash,
are requested during the test, so it would be nice if an HTTP Proxy
could be
used to create the initial test plans by recording actual browser
sessions.
Some things I’ve looked at so far:
JMeter:
seems difficult to make dynamic test plans here that will modify
requests
based on earlier responses
test plans are in XML, although I suppose I could use a ruby DSL to
generate the XML, but since the proxy records to XML I’d also have to
generate ruby from XML
The Grinder:
Looks awesome, and I may end up going with this, but it would be nicer
to
use ruby instead of Jython in order to leverage my teams existing code
and
skills.
shell/curl/ruby script mish mash:
A lot of control, and curl handles cookies, but no real framework
support
and no proxy recording
Ruby code:
Doesn’t really seem to be any test framework that deals with
concurrent
requests like this
I assume JRuby would be required (and would be sufficient) to get real
threading
I feel like there are probably a lot of pieces that could be tied
together
to make what I want, but unsure of what they are or which ones to
concentrate on. Hpricot? Mechanize?
Anyone have any ideas? Is anyone doing load testing via ruby?
I’ve made similar searches and come up with the same emptiness. There
are two major commercial load testing tools, LoadRunner and
SilkPerformer. I have about seven years of experience with
SilkPerformer and almost none with LoadRunner. There are a few open
source recorders, but none I have found that are anywhere near as good
as SilkPerformer.
What you could do is get a low-end SilkPerformer or LoadRunner
license … just enough virtual users so you’d get a recording
capability … and then write a Ruby run-time and parser to translate
the “native” language to Ruby. But check the EULA to see if this is
legal. I never tried it because I consider it a waste of effort.
Really, you need to make a business decision here. What is the cost to
your business of the time you spend building a load testing framework
when you could be developing and testing your product? I’d recommend
you seriously consider buying a license for LoadRunner or
SilkPerformer if performance of your application is mission-critical.
On Wed, Feb 25, 2009 at 12:58 PM, M. Edward (Ed) Borasky [email protected]wrote:
In other words, I’d like to run say 50 concurrent user sessions where
use some of the same DSL from my selenium or integration test helper
images/css/js/flash,
based on earlier responses
I feel like there are probably a lot of pieces that could be tied
I’ve never met a happy clam. In fact, most of them were pretty steamed.
Thanks Ed,
Good to know I wasn’t missing something obvious.
I’ve contacted sales reps for SilkPerformer and LoadRunner. My fear (and
unfortunate experience) with this kind of software is that you spend a
lot
to get a complicated but powerful system that then requires more time
(and
money) to make actually work. They both seem to have a lot of features
that
I don’t need, so I was hoping to find something more streamlined and
ruby-oriented first. It’s good to be aware of all the alternatives
though,
so thanks for the pointer. I had heard of LoadRunner but not
SilkPerformer.
so thanks for the pointer. I had heard of LoadRunner but not SilkPerformer.
50 concurrent virtual users probably won’t cost you a lot. They usually
charge a base price plus a price per virtual user. The assumption is
that your business will grow, you’ll need to emulate more users, and
you’ll be able to buy more virtual users.
For what it’s worth, the real “secret sauce” in the recorders is the
ability to extract session data, cookie send / receive, parsing data out
of the returned HTML and branching on it, etc. Pure record / playback
isn’t all that hard to do, it’s the “other stuff” that makes these tools
worth the money.
P.S.: It warms my heart to know that there are actually web application
developers out there who realize the need for load testing. There are
some big name web sites out there that seem to go into deep space when
they update the database, for example. And load testing can also give
you good news … you may need less server than you’re planning on.
Its easy to distribute, rather easy to configure (in XML…), is able
to handle dynamic stuff like “create a resource, parse the id out of
the answer and create a child resource” rather easily. There is also
an easy way of consuming the answer using an Erlang function.[1] The
Documentation is quite thorough.
It also provides a recording proxy and AFAIK there is a parser for
apache logs.
I also have a small sinatra-app serving the logs that I want to
release next week or so.
seems difficult to make dynamic test plans here that will modify requests
based on earlier responses
test plans are in XML, although I suppose I could use a ruby DSL to
generate the XML, but since the proxy records to XML I’d also have to
generate ruby from XML
At first it looks rather difficult to make dynamic tests in JMeter,
but it’s really not hard once you get into it. The caveat is, you use
javascript & regexes & xml or CSV files. I used to do this frequently
at my old job.
Under no circumstances even bother with TestComplete by AutomatedQA.
It’s a piece of junk.
Unless you have the cash to spend on expensive software and a few
training classes for it, I would delve further into JMeter. Record
snippets of what you want to do into multiple test groups, trim down
the calls to just what you need, code from there. You may want to
start following their mailing list. It helped me a ton.
It really depends on why you want to do this. If you simply want to
push a workload that is representative but without any concern about
absolute correctness then a JMeter or LoadRunner or Grinder can be
good. LoadRunner has a great UI. This allows you to stress test and
see where the hotspots are. New Relic is an excellent complement to
any of these.
If you want a realistic load test and want to actually estimate
absolute capacity then you’re choosing to do something harder. The
best tool is the open source httperf. Its the only tool here that uses
an open workload model. You will find that LoadRunner and the other
tools I suggested will always overestimate system capacity. That is
because in these tools, a virtual agent won’t fire a new user request
until the prior one completes.
This of course is not what happens in real life, where a heavy tailed
Pareto distribution of requests is typical. Httperf addresses this
issue and allows you to get data that is more realistic in an absolute
sense.
If you’re wanting to tune and improve performance or throughput and
not simply measure it that’s a whole other ball-game. Be prepared to
take some time and learn, or hire some help.
Peter B.
917 445 5663
PS - If anyone needs assistance with tuning Rails apps, website
capacity planning, load testing, please ping me. I am available. I
work at the application code, web server, DB server, operating
system, TCP stack, physical network and hardware layers.