> On Tue, Feb 03, 1998 at 04:55:13PM -0900, Todd Nagengast - Systems Wrangler wrote:
> > I can't speak for any of the other methods with empirical data, but I can
> > tell you that for a true split the bits bandwidth type of load sharing,
> > eBGP doesn't seem to do a very good job. We're currently trying to do it
> > with UUnet over two FR T1s. The routes will get cached on the FR subinterface
> > and then all traffic for the cached route will go out that interface.
>
> A mistake often made in looking at various way to load share/balance
> parallel links is that people assume that it's a routing protocol feature.
>
> While modern routing protocols do support _routing_ over equal cost
> paths, it's the forwarding mechanism that determines the load sharing.
>
> The per-packet load sharing done in D/CEF is probably what folks are
> looking for, but it has a draw back in the sense that it can cause out of
> order packet arrivals and compromise the end to end performance.
>
> -dorian
Dorian,
I agree that terminology is confusing sometimes. How about 'route
balancing' or 'prefix
balancing' for a routing protocols which distributes its routes among
known equal cost paths. Then for the forwarding we can accept 'traffic
balancing' or 'load balancing' refering to
the true distribution of bytes and packets.
As for traffic balancing there might be different methods such as round
robin, destination hash,
source-destination hash and even using port numbers. While round robin
will more likely to cause
out of order scenario, source-destination hash might be better choice.
$0.02
Greg
This archive was generated by hypermail 2b29 : Sun Aug 04 2002 - 04:13:15 EDT