Strange loadbalancing behavior

Marko Leopold · ‎01-27-2010

Hello!

In out testing enviroment is a strange loadbalancing behavior. because of the migration from iPlanet-Webserver to Apache there are currently some tests with 1 Apache and 2 iPlanet servers. the servers have the same hardware. the operation system on iplanet is SLES 9 and for apache there was an upgrade to SLES 10. We use a CSS 11506 running 8.20.4.02 which is the latest firmware. the way the packets are running from client to CSS. from there to some web accelerators and back to CSS. from CSS to the webservers. there are 3 services running on the servers which are configured in the services section. the difference is in the destination port. configuration for the web accelerators is

service aclxxx_way[1/2/3]

ip address xxxx

protocol tcp

port xxxx

active

the webserver are configured

service webxxx

ip address xxxx

protocol tcp

port xxxx

active

The contents are all configured with no persistence and loadbalancing round-robin on all webservers.

During the tests we see that the iPlanets are getting a lot more of the traffic then the Apache. the load (response time) of the Apache is quite low and mostly 5 times lower than the iPlanets. Connections is the same. It seems like the Apache is getting much much less work then the iPlanets. Does anyone have explanation for this? Do you need more information?

Marko

jason.espino · ‎01-27-2010

Marko,

The load balancing behavior your experiencing is indeed odd. Can you please provide the output of the content rule configuration the services are applied to that you are noticing these issues? Do you know if the Apache server is having issues properly terminating connections? Is there a lot of connections in a TIME_WAIT state on the Apache server? Do you know if any of the servers have HTTP keep-alives enabled? Has the Apache server experienced more state transitions then the other servers?

If you can provide more information that would definitly be helpful.

Thank you,

Jason Espino

Marko Leopold · ‎02-01-2010

Hello Jason!

Okay, this was a lot of questions. i will try to answer you as much as i can. The rest i will ask the server administrators.

The content of the servers is configured like this

content Webserver_way1/way2/way3
    flow-timeout-multiplier 38
    vip address xxx
    port 8090/8091/8080
    protocol tcp
    primarySorryServer mtc_ext
    add service webserver1
    add service webserver2
    add service webserver3
    active

and the contents for the web accelerators are configured like this

content Accelerator_way1/way2/way3
    flow-timeout-multiplier 38
    vip address xxx
    port 8081
    protocol tcp
    url "way1/way2/way3"
    no persistent
    primarySorryServer mtc_ext
    add service acl01_ext_e
    add service acl02_ext_e
    add service acl03_ext_e
    active

Okay, thats for the CSS configuration. The issue is noticed at the webserver-content.

The questions for the Apache i will ask our server administrators, but i can tell the Apache is not experiencing any state transitions at all. running very smoothly.

i hope this informations can help you understanding the problem.

Marko

Gilles Dufour · ‎02-02-2010

You didn´t give us the most important information .... the content rule.

There we can see if you have stickyness configured ... most probably the case...

A loadbalancer can´t guarantee equal load on each server when stickyness is configured.

Regards.

Gilles