[c-nsp] per flow load balancing

Everton da Silva Marques everton at lab.ipaccess.diveo.net.br
Wed Dec 28 09:36:30 EST 2005


Rodney,

On Tue, Dec 27, 2005 at 02:22:08PM -0500, Rodney Dunn wrote:
> On a 65xx/76xx the flexwan's don't do any packet forwarding.
> 
> It's all hardware based as long as you don't have a feature
> on that causes the traffic to be punted.

Many thanks for the clarification. Then I assume the
forwarding is performed by the PFC. In such case, there is
one small doubt that I hope you could bring light upon: I
can't tell whether the PFC's hardware-based load-sharing
would ever feed L4 ports into the hashing.

So far, this is what my research has revealed:

The document below seems to imply (?) that hardware
load-sharing for MPLS-labeled packets hashes the output
interface from exactly the source-destination pair:
  
  "If the packet has three labels or less and the underlying
  packet is IPv4, then the PFC3BXL or PFC3B uses the source
  and destination IPv4 address." -- http://tinyurl.com/clq35

That is, under 7600, hardware-based load-sharing for labeled
packets doesn't seem to consider L4 ports.

BTW, in the same scenario, do hardware-based load-sharing
for pure IPv4 packets use L4 ports in the hashing? I have
found the reference below suggesting (?) that L4 ports don't
affect the hashing for plain IPv4 packets:

  "With a PFC (and DFCs, if present), hardware Layer 3
  switching uses per-flow load balancing based on IP source
  and destination addresses." -- http://tinyurl.com/bb89s

Please let me know whether the PFC should be able to
actually use L4 ports for hardware-based per-flow
load-sharing.

Thanks again,
Everton


More information about the cisco-nsp mailing list