[c-nsp] IP Cef load sharing, quick question

[Peter Rathlev]

On Fri, 2008-11-07 at 08:28 -0500, Drew Weaver wrote:
> We have a Simple L3 switch (I think it's a 2960G) that we need to do
> some even simpler fault tolerance and load sharing on.

They're not 2960s, since those are L2 only. Maybe 3560s?

> We were going to connect this switch to 3x switches upstream and then
> do something like this:
> 
> ip route 0.0.0.0 0.0.0.0 g0/32 gwip
> ip route 0.0.0.0 0.0.0.0 g0/33 gwip
> ip route 0.0.0.0 0.0.0.0 g0/34 gwip

As Steinar mentions, you should use a specific next hop address. I
assume that the three interfaces are routed ports, or that they use
seperate VLANs. What's upstream?

> When we were testing we noticed some (well, quite a bit) of strangeness
> with traceroutes and the like (many multiple hops for the same, hop..
> etc)

What exactly do you mean with "many multiple hops"? Different answers
for multiple requests with the same TTL (same hop in traceroute) is not
all that unnormal for multiple paths -- each path is elegible for the
traffic, so each next hop router can answer.

The L3 switches (i.e. not software based routers) typically use a hashed
load sharing algorithm, resulting in per destination or per source (or a
combination) load sharing. They could also include the ports in the
hashing, meaning that a traceroute using a different source port per
probe would result in different next hops. If this is unwanted, you can
change the algorithm to something that doesn't include L4 ports.

> is there a better way to do what we're trying to achieve?
> 
> We were thinking about maybe doing VRRP on the 3 switches upstream but
> then we would only be using 1Gbps and the goal is to be able to use "a
> little more than" 1Gbps.

You _could_ use GLBP as a load sharing enabled equivalent of VRRP. Don't
know is your hardware/software supports it though. And equal cost
multipath (ECMP) would be my preferred choice if possible.

Regards,
Peter