Load Testing Help - how to improve perfromance

[Baptista, Paulo]

Thanks all for the feedback.  Few more notes to report.

Our IDP servers were running at about 60-70% overall cpu utilization.  The
IDP Java proc itself was using about 125% of cpu (we have 2 cpus)

We adjusted the way we logged to INFO and WARN, from ALL and DEBUG.  Doing
that dramatically improved performance.  We went from 4000 logins in 5
minutes to 17000, with 1000 concurrency.

Reason why we want to test with high concurrency is because during our
Admission Decision night, we have about 30k kids wanting to see if they got
into the University.  We're planning to go higher! Wish us luck! ;)

Thanks all,

Paulo



On Tue, Nov 11, 2014 at 12:52 PM, Cantor, Scott <cantor.2 at osu.edu> wrote:

> On 11/11/14, 4:17 PM, "Baptista, Paulo" <paulo_baptista at brown.edu> wrote:
>
>
>
> >How are all of you?  We're running some load tests against our QA Shib
> >IDP Infrastructure.  I'd like to share our current setup and results, and
> >hopefully get from the community some tips on how to improve performance.
>
> First you have to figure out the bottleneck and then determine if it's
> fixable. Chances are it's not, and you simply need more capacity. If it's
> CPU bound, then you need more CPU, that's it. It scales linearly.
>
> >2 Red Hat servers, IDP load balanced behind F5
> >each with 2CPU, 2.5Ghz
> >8GB ram
> >java 1.6.0_14
> >-Xms256m -Xmx6g -XX:MaxPermSize=1g -XX:-DisableExplicitGC
> >-XX:+UseParallelOldGC
>
> Java 1.6 is dead, so I'd start by using one that isn't, since it's newer
> and might have performance improvements anyway.
>
> >We're running load test with 1000 concurrent users for 5 minutes. We get
> >about 4000 successful logins. We're using Jmeter.
>
> I seriously doubt that's a realistic load, but if you think you need that
> and it's CPU bound, then I guess you need more cores. Pretty simple. If
> it's not CPU bound, then the IdP isn't the bottleneck.
>
> -- Scott
>
> --
> To unsubscribe from this list send an email to
> users-unsubscribe at shibboleth.net
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://shibboleth.net/pipermail/users/attachments/20141111/4c0fd577/attachment.html