graeme at graemef
May 9, 2008, 6:05 AM
Post #3 of 3
(327 views)
Permalink
|
|
Re: [lvs-users] active-active by a multipath routes
[In reply to]
|
|
On Fri, 2008-05-09 at 12:06 +0300, Siim Põder wrote:
> Is anyone running active-active(-active...) LVS setups?
I'm not (yet) but I may be doing so sometime this year. Current thinking
is (per your email) to run Quagga, Zebra, or ospfd (or something else)
on the directors themselves which will announce the VIPs into the local
network.
The local network devices will then work out the best path for traffic
flowing through them; OSPF is designed around a cost and hop-count model
so having multiple routes originated in different places should mean
"closest network wins" from a client perspective - although this is
untested!
The interesting part is how you make sure the traffic returns to the
clients. In the case of -DR this isn't really a problem, but using -NAT
could be difficult if the realservers can receive, and return, traffic
to either director.
By extension, you could split a cluster into two halves and have each in
a different physical and logical location using this model, but that
adds complexity at the backend if you're sharing file data and/or
databases. What that would give you is, for example, proper geographical
resilience such that if one location loses power, hey? Who cares? We'll
just talk to the other one instead!
And as an added bonus, if you have multiple equal-cost paths to
directors in the same location, OSPF can be made to (crudely, in some
cases) load-balance between them. This gives a bit more even-handed
loading but does add an extra thing to go wrong.
Graeme
_______________________________________________
LinuxVirtualServer.org mailing list - lvs-users[at]LinuxVirtualServer.org
Send requests to lvs-users-request[at]LinuxVirtualServer.org
or go to http://lists.graemef.net/mailman/listinfo/lvs-users
|
