|
On Tue, 2007-12-18 at 13:23 +1100, Adam Niedzwiedzki wrote:
> I'm running (LVS via keepalived) and it's humming along nicely.
Good stuff.
> I have a couple of clients that are just using wrr and rr for the scheduling
> across their boxes.
> I have another client that has requested just a basic failover solution,
> they don't want to load balance across their serves they just want failover.
Right...
> How do I setup LVS (or the config in keepalived) to do this. Send all
> requests to one server, unless it disappears then failover to the other, on
> it's return fail back.
[see the keepalived.conf man page, or the SYNOPSIS included in the docs
for terminology below if you haven't seen it before]
There are several ways to achieve this. The most basic is to setup the
virtual_server instance with only a single real_server, and then define
a sorry_server which will take over when the single real_server fails.
The issue with this is that the sorry_server isn't healthchecked, as it
isn't part of the normal topology. You have to guarantee absolutely that
it will be (a) up, and (b) able to handle your load upon failover.
The second, slightly more complex method is to define two real_servers,
but make one have a weight of (for example the maximum) 65535, the other
a weight of 1. Obviously with WRR this means 1 in 65535 requests gets
handled by your "spare" server; whether your users would accept that is
up to them.
The third method would be to use a combination of weight, the wrr
scheduler, persistence, and the sysctls controlling the behaviour of the
persistence templates when a real_server is removed or becomes
quiescent. These are:
/proc/sys/net/ipv4/vs/expire_nodest_conn
/proc/sys/net/ipv4/vs/expire_quiescent_template
The first expires all entries for a given real_server when it is removed
from the pool; the second removes all entries for a given real_server
when it becomes quiescent (ie. weight becomes 0).
I hope that helps.
Graeme
|
