[c-nsp] 3750 etherchannel only using 1 port

[Andre Beck]

Hi Fred,

On Wed, May 07, 2008 at 10:29:21AM -0400, Fred Reimer wrote:
> I'm not sure what you meant by the term "purely statistical" when referring
> to EtherChannel load balancing, but I think it may give a false impression.

I'm used to this term meaning "there is no load balancing unless a large
number of source and/or destination systems are involved, so that every
single deterministic decision will lead to a statistical distribution".
Maybe it's not the correct term or does mean something different in
english (which isn't my native language), so if you have a better term
that states this in a single word, I'm all ears ;)

> If you meant that the results of the load balancing would be "by chance" and
> statistically average out over time then that's incorrect.

Not over time. Over sources/destinations. Lot's of them.

> EtherChannel
> load balancing is very deterministic.  You can tell exactly what port will
> be selected based on the load-balancing protocol and the number of ports.

Yep, I know. In my statement

>> Of course you will see a certain flow always taking just one member
>> egress path, any load balancing is purely statistical.

I hoped that the first part of that sentence makes this clear. Of course
it all depends on your definition of "flow", which here is essentially the
load-balance algorithm, like "Src XOR Dst IP" (somehow hashed to the
number of members).

I should have pointed out that the most relevant thing to know here is
that on L2, there is *never* per-frame load balancing, at least not in
a typical switch. People coding operating systems sometimes don't see
all the potential problems that are involved and simply provide such
means anyway (and they often even work).

> There's even a SP command to tell you what would be selected on the 6500
> platform (test etherchannel load-balance interface pox ip y.y.y.y z.z.z.z in
> a remote login switch session).

Cool, this works well on my 3560Es too. It even tells me that indeed, a
changing source IP address leads to changing egress members on my L2
port-channels with src-dst-ip load-balancing. So it does work as expected.
Shouldn't have been so doubtful here ;)
 
> If you have a relatively low number of very high throughput streams you can
> use the actual protocols to map out what ports would be selected, and choose
> the appropriate protocol for your specific needs.

As long as your hardware supports such balancing algorithms (up to
transport layer port numers etc). This is no option with the hardware
in question:

sw-ibh-xg1(config)#port-channel load-balance ?
  dst-ip       Dst IP Addr
  dst-mac      Dst Mac Addr
  src-dst-ip   Src XOR Dst IP Addr
  src-dst-mac  Src XOR Dst Mac Addr
  src-ip       Src IP Addr
  src-mac      Src Mac Addr

So the best you can do in an IP shop is src-dst-ip so at least you
balance on more than the router MACs (which in the worst case leads
to no balancing at all).

Andre.
-- 
   Real men don't make backups of their mail. They just send it out
    on the Internet and let the secret services do the hard work.

-> Andre Beck    +++ ABP-RIPE +++      IBH IT-Service GmbH, Dresden <-