Forum Discussion

0veracl0ud_1734's avatar
0veracl0ud_1734
Icon for Nimbostratus rankNimbostratus
Feb 08, 2016

Webapp loadbalancing issue

Hello,

 

I'm facing a loadbalancing issue. Here the context : I've to loadbalance a web app (port dest 80) which the client requests are only XML files sent to the server through the POST METHOD. The server respond back with an XML files also.

 

Client --> F5 --> Pool --> Node 1 or 2

 

To do so, I configured a VS with an ASM profile (to validate XML scheme) with no persitence (the value is set to None) and no OneConnect profile. On the pool level, I configured it with the Round Robin Method and two nodes.

 

To make my tests, I got an "homemade" XML requester that can send several XML requests to the servers. It's able to handle multiple connections. Each connection start with a keep-alive information in the HTTP header.

 

My issue : when I forge a load test with multiple connections (2 or 4), I observed that all connections (via pool stats) go to only one node instead of loadbalance on both... But sometime it works and I don't know why.

 

Do you have an idea on what am I doing wrong ?

 

Thanks a lot for your help :)

 

4 Replies

  • Hi 0veracl0ud,

    for OneConnect disabled virtual servers, you would still maintain a sticky connection to the selected node, till your client side connection is getting closed. The LB would only be performed for every new client connection...

    To get a new load-balancing connection on each single HTTP request, you may have to turn off HTTP keep-alives in your homemade XML requester, or you may want to manually

    [LB::detach]
    the backend connection on every consecutive request.

    when HTTP_REQUEST {
        LB::detach
    }
    

    Cheers, Kai

  • Hi Kai,

     

    Thank you for your answer.

     

    Actually, I would like to keep a HTTP connection (with its several XML requests) on a node. But if I have two (or more) HTTP connection, I would like to loadbalance on both nodes. Currently, when I made 2 connections, both go to the same node... and sometimes (but not often) on both nodes.

     

  • Thanks Kai.

     

    I will try with this method. Before sending a post here, I tried with the Ratio (member) method with a ratio of 1 on each node. But I didn't see any change.

     

  • Ok, thanks a lot for these informations. Where did you find as accurate specification about the LB methods ? I'm very interesting :) In the offical documentation I didn't find any info about statistical decision, etc.