Hi Rainer,
Thanks for the well written response and some ideas to
get past the problem.
I agree the 4ms might be latency but until I can
measure at Apache web server or Tomcat more acurately
I will wait to say what it really is.
I am using %D and as you said the timer is not very
accurate. Hopefully I will switch to linux soon and
the whole measurement question will be moot.
Regarding paralallelism to me the best you can
possibly ever get is the response time for a single
request then after that all the measuements just get
worse. But if I can get a transaction under a
microsecond and even if I had only one core then
serially I can do over a million transactions per
second and maybe with two cores 1.5 million per
second.
As I mentioned in another email I just sent one most
my transactions because of the overall design are
under 1500 bytes or 1 ethernet packet.
I will try some of your suggestions and others brought
up by the email group so guess I will be busy tonight.
Thanks,
-Tony
--- Rainer Jung <[EMAIL PROTECTED]> wrote:
> First: measurement:
>
> Apache access log can use "%D" in the LogFormat,
> which for Apache httpd
> 2.0 and 2.2 gives microseconds response time.
> Obviously the real
> resolution depends on the operating system. The
> 15msec you mentioned is
> way to long for a OS timer resultion.
>
> mod_jk provides an apache "note" named
> JK_REQUEST_DURATION, which you
> can include as "%{JK_REQUEST_DURATION}n" again in
> the access log of
> Apache httpd. This measures Tomcat overall response
> time from the point
> of view of mod_jk (microseconds formatted as
> seconds.micro).
>
> Finally Tomcat itself can log milliseconds via "%D"
> in his access log.
>
> But: when you talk about 4ms transaction time, it's
> more about latency,
> which is only one aspect of performance. Usually
> latency is only
> important if it gets to big, but decreasing latency
> below the OK level
> is not very important. If latency is below an OK
> level, troughput is
> more important wrt. performance.
>
> throuput * response time = parallelism
>
> In general if you want to measure throughput in a
> real life throughput
> driven system, you need to increase the client
> parallelism to find out
> how much throughput you can achieve. Response time
> (latency) will go up
> too, but in some parallelism range, throughput will
> increase more.
>
> If you add a component like httpd/mod_jk to the
> system, it will increase
> latency. If your transactions are very fast, the
> added latency might be
> relatively huge. In terms of throughput this will
> usually not produce a
> penalty, and in some cases, if you can effectively
> use the features of
> the added component, throughput might go up.
>
> Regards,
>
> Rainer
>
> Tony Anecito wrote:
> > Hi All,
> >
> > As I mentioned in my previous email I was able to
> get
> > my system performance using Apache-Tomcat-Jre1.6
> such
> > that I was getting 4msec transaction response
> times.
> >
> > My problem is I can not tell how much of that
> response
> > time is spent in tomcat versus Apache. I am using
> > windows 2000 where the configuration is Apache -
> > (Tomcat embedded in JBoss 4.2.1) using Mod_jk. The
> > problem I can measure the transaction in the
> servlet
> > which is in the sub-microsecond range but at
> apache
> > access logs or in the mod_jk logs I am seeing
> either 0
> > or 15msec. I suspect that is due to the timers
> used by
> > Apache.
> > I would like to know how long Tomcat is taking for
> the
> > request. I am using web services inside of Tomcat
> so
> > the request goes through that layer.
> >
> > Also, can anyone tell me the performance
> differences
> > they have see when Tomcat is running on Linux
> versus
> > Windows? I have heard anything between 1 to 30%
> > improvment.
> >
> > Also, any hints on how to improve Tomcat
> performance
> > appreciated.
> >
> > Regards,
> > Tony
>
>
---------------------------------------------------------------------
> To start a new topic, e-mail:
> [email protected]
> To unsubscribe, e-mail:
> [EMAIL PROTECTED]
> For additional commands, e-mail:
> [EMAIL PROTECTED]
>
>
____________________________________________________________________________________
Check out the hottest 2008 models today at Yahoo! Autos.
http://autos.yahoo.com/new_cars.html
---------------------------------------------------------------------
To start a new topic, e-mail: [email protected]
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
