[PLUG] benchmarking webserver performance ?

Kurt Sussman plug at merlot.com
Wed Mar 19 04:46:20 UTC 2008


M. Edward (Ed) Borasky (znmeb at cesmail.net) typed this ...
> Russ Gilman-Hunt wrote:
> > I've gotten my boss used to the idea of "benchmarks" so that he can not
> > only ask "is the load high on serverX" but also "how does the load today
> > compare against last week's load?"  ( good boss, have a donut ). 

That's one useful metric. I try not to confuse the PHBs with too many
metrics, but you might enjoy others.

> > Now he wants me to look at a webpage on serverX and gauge the response.
> > On one hand, it responds quick, but on the other several included bits
> > ( you know, graphics, javascript libraries, all that web2.0 crap ) can
> > take seconds to materialize. 

You can get this from your logs.

What is the purpose of this measurement? Log analysis will tell you
more than just page load time. 

> Question: has your boss given you a budget, or does he want you to build 
> this yourself? The reason I ask is that there are quite a few good load 
> testing tools out there. The two best IMHO are Borland SilkPerformer 
> (which I know well) and HP/Mercury LoadRunner (which I've never 
> touched). These are the expensive ones. There are also open source tools 
> and less expensive commercial tools, but I'd recommend investigating the 
> "name brand" tools first.

There are a ton of tools that can do load, stress, and performance
testing.  Silk is fine if you have bucks to spare and time to write the
scripts.  I haven't worked with it since 1999, and I'm sure it has
changed. But I also worked with it when it was in beta (1994) and there
was not a big delta to 1999.

> As far as knowing when "the last piece is done", that's not easy for a 
> variety of reasons. Essentially you need to get fairly intimate with the 
> browser, and it's highly dependent on how fast the client PC is, how 
> much RAM it has, and what else it happens to be doing.

You could actually script a real browser (Selenium, Watir), but it would
have to be on a network that is representative of the networks your
customer/clients are using. Better to use the data already being
collected in the log.

How about Cricket or if you don't want to use the logs for some reason,
Jmeter or httperf or something from this page:

  http://www.softwareqatest.com/qatweb1.html#LOAD

--Kurt
-- 
----------------------------------------------------------------------
     Merlot Research Group, Inc               kls[at]merlot.com
    "Quality is free, but only to those who are willing to pay 
     heavily for it." (Lister, DeMarco: "Peopleware")



More information about the PLUG mailing list