|
Abstract:
I'm going to favor RIP and say why, then I'm going to favor OSPF and say
why... :) Then I'm going to summarize.
> To my mind the idea of using a routing protocol would be best if a routing
> protocol is already used on the network. After all if you are using fwmark
> to manage CIDR networks that are routed to a Linux Director managing the
> path that this traffic takes really is a routing problem and it would make
> sense to make use of any existing routing infrastructure. Though would be
> thinking more in terms of an IGP such as OSPF rather than an EGP such as
> BGP, though this is really an implementation issue.
I agree completely -- it makes sense to use an existing routing protocol. I
also agree withat using an IGP rather than an EGP would be better, which is
why I originally suggested RIP rather than BGP. Routing CIDR networks could
be done with RIP2. (RIP1 doesn't support CIDR) The only reason I suggested
RIP over OSPF was just because it is very simple to setup and might be more
appropriate for LVS as OSPF is more for larger networks that need a lot more
organization than two LVSs and a few IPs.
Like RFC1723 says:
=====
With the advent of OSPF and IS-IS, there are those who believe that
RIP is obsolete. While it is true that the newer IGP routing
protocols are far superior to RIP, RIP does have some advantages.
Primarily, in a small network, RIP has very little overhead in terms
of bandwidth used and configuration and management time. RIP is also
very easy to implement, especially in relation to the newer IGPs.
=====
And personally, I'd be afraid of having to deal with managing OSPF for LVS
clusters combined with internal OSPF on a network. I'm not sure how much I'd
want advertised where... so I'd have to filter certain things off.
HOWEVER... the timeout issue....
> The question that I have is - and this shows where my knowledge of routing
> protocols runs out - what is the relative cost of waiting for various
> routing protocols to converts as opposed to using a fail-over mechanism
such
> as heartbeat or your etherbeat?
Having a heartbeat-like system is nice because you can tweak the failover
timeouts and sensitivity.
It takes 180 seconds for RIP2 to expire a route (unless the particular
router sending the updates sets its metric to 16, but in a failover
situation this probably won't happen). This has always been one of the
downsides of rip -- slow convergence.
OSPF2 allows more flexibility which might actually make it a good
replacement for heartbeat. Adjusting these configuration items:
HelloInterval
The length of time, in seconds, between the Hello packets that the
router sends on the interface. Advertised in Hello packets sent
out this interface.
RouterDeadInterval
The number of seconds before the router's neighbors will declare
it down, when they stop hearing the router's Hello Packets.
Advertised in Hello packets sent out this interface.
Hello Timer
An interval timer that causes the interface to send a Hello
packet. This timer fires every HelloInterval seconds. Note that
on non-broadcast networks a separate Hello packet is sent to each
qualified neighbor.
Might allow for OSPF to update quick enough. On a small network, OSPF
wouldn't send a great deal of information with each advertisement. Plus,
OSPF adds an extra layer of security.
So basically...
*) Things work the way they are currently setup... hacked as it might be.
*) Changing to an IGP might allow for more than two LinuxDirectors and would
use an existing dynamic routing protocol to do ... well ... dynamic routing.
:)
*) RIP is very simple and would be very easy to setup... but it would take
180 seconds for a failover.
*) OSPF is more complex -- perhaps a little too complex for something so
simple as this -- but it has a great deal of flexibility.
Now what do you think about all of this? Do you think that RIP2/OSPF2 could
be useful for LVS? Would we gain a great deal over our current methods of
failovers?
ANOTHER THING -- Weren't we at one time talking about using something like
heartbeat to advertise persistence information so that when one LVS died,
the one that took over would continue routing exactly the same way causing
virtually no donwtime for anyone? Using an IGP to handle these sort of
failovers wouldn't share this information... However -- back in the day when
we were talking about this I said that using an IGP as a *MODEL* for such an
advertisement protocol might be a decent idea. For example, we could send
advertisments like OSPF's LSA's and even offer authentication like OSPF. We
could use broadcast or multicasts to send this information... etc...
And perhaps this LVS-IGP could advertise this info as well as typical IGP
info so that routers surrounding LVS would still route correctly; the LVS's
would be the only ones that cared about the other information.
Of course that's quite a project.
I hope that inspires some imagination.
All the best --
Ted
|
