[e2e] Link Aggregation

[Cannara]

Valentin, what you (and I) want is essentially doable now in Sonet-like
systems, with "Virtual Concatenation" -- meaning links can be dynamically
sized (in some multiples of, say, T1) to take advantage of unused bits in
synchronous framing (see AMCC, Vitesse...).  Unfortunately, the IEEE standards
for bridging seem to preclude DLC link aggregation per flow.  Routers can, of
course, be configured to do this (e.g., Cisco's "Fair" vs "Round-Robin"
parallel route-path use) with, say, multiple T1s, but that capability
disappears at the outer LAN interfaces.

Alex

Valentin Ossman wrote:
> 
> The reason that I was looking for a Link Aggregation method in the first
> place was that there is no way to establish a 2Gbps connection between 2
> stations using 2 GbE lines.
> Link aggregation does not allow higher bandwidth per connection, but
> only high bandwidth aggregated traffic with the limitation that the
> maximum connection bandwidth is 1Gbps.
> 
> Still, does anyone know of any effort to define a link aggregation
> method that will permit high bandwidth connections? This will be a very
> useful method to increase bandwidth.
> 
> Valentin
> 
> > Order management is indeed a feature of available network processors
> used
> > in
> > newer routers & similar forwarding devices.  Incoming pkts are, at the
> > very
> > least, identified by source port (channel on a multiplexed port) and
> time
> > stamp, and will not be sent from an output queue (or queues) until all
> > pkts
> > that had arrived before, with the same identifier, have been sent.
> This
> > was
> > common in the design of ASICs for routers and was carried on into
> network-
> > processor chips, either in hardware dedicated to order management
> (IBM,
> > Vitesse...), or firmware (Intel...).
> >
> > IEEE link aggregation at Layer 2 (DLC), by the way, has been
> incorporated
> > into
> > the design of single-chip switches/bridges, but brings with it some
> > interesting shortcomings.  One is that the IEEE standard did not allow
> > multiple paths (links) to be used for any one DLC association.  Thus,
> if
> > one
> > has a system capable of filling more than one link's capacity to
> another
> > system, link aggregation does no good -- pkts between the two systems
> are
> > passed on one link only, as identified in the bridging table entry for
> the
> > two
> > DLC ends.  Aggregation only helps when there's a mix of DLC
> associations.
> >
> > Alex
> >
> > Alia Atlas wrote:
> > >
> > > I expect that you may run into problems with one TCP connection
> because
> > > most routers try to avoid misordering packets of a single microflow.
> > Thus,
> > > even though there may be several interfaces which are part of an
> > ethernet
> > > link aggregation, the traffic from a single TCP connection would
> > probably
> > > be directed towards only one of those interfaces (for a good
> > implementation
> > > which is trying to avoid misordering).
> > >
> > > Alia Atlas
> > >
> > > At 09:56 AM 5/13/2003 -0400, RJ Atkinson wrote:
> > >
> > > >On Tuesday, May 13, 2003, at 08:59 America/Montreal, Valentin
> Ossman
> > wrote:
> > > >>Does any one know a link aggregation method for several network
> > adapters
> > > >>(let's say 4 Gigabit ports) in a way that it will be possible to
> > achieve
> > > >>one high bandwidth (4Gbps) TCP connection?
> > > >
> > > >IEEE have a published standard for Ethernet Link Aggregation.  I'd
> > > >start by looking that up.  Not clear to me that this relates to the
> > > >charter of this mailing list however, so followups probably belong
> > > >elsewhere.
> > > >
> > > >Ran