[lvs-users] LVS RR becomes unbalanced after time

Graeme Fowler graeme at graemef.net
Wed Jul 6 08:54:42 BST 2011


On Tue, 2011-07-05 at 09:37 +0200, Luc van Donkersgoed wrote:
> I don't see any mention of segfaults in my logs, but did manage to solve my problem. Apparently I forgot some basic configuration on my webservers, which screwed up ARP. I don't understand the specifics yet - I will delve into that today - but the solution was to add these lines to /etc/sysctl.conf:

This is a very well-understood and documented issue but one which
catches out a lot of people. Look in the HOWTO, Google or search the
archives for "the ARP problem".

In short: For LVS-DR to work, a realserver has the VIP bound to a
"hidden" (dummy, loopback or other) interface. This means when the
upstream router from the directors sends an ARP request saying "hey
guys, who's got the VIP?" one of two things can happen:

1. The active director replies first; this is good and expected, and all
works normally.
2. A realserver replies first; this is not so good, and means all
requests bypass the director and go to the realserver directly.
Everything becomes unbalanced.

The sysctls you set prevent the realserver from replying to ARP requests
on for addresses bound to interfaces which are supposed to remain
"hidden".

Graeme





More information about the lvs-users mailing list