Tracy R Reed wrote:
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
Andrew Lentvorski wrote:
The problem is that Ethernet uses CSMA/CD (carrier sense multiple
access/collision detect) and cannot, by its nature, provide transmission
guarantees.
Only on non-switched networks right? You generally don't have collisions
on switched networks which is why you can do things like use ethernet as
an ATA bus with AoE.
You don't have CSMA/CD collisions, but you still have some arbitration
when two different ports (A & B) need to connect to the same third port
(C). The switch has to arbitrate.
At this point, you have problems. The switch doesn't *have* to release
a connection between A and C so B can talk. Most do, but that is
neither guaranteed nor specified.
I'm pretty sure that AoE wants switch level flow control. There's your
arbitration and there goes any performance guarantees. Also, I'm pretty
sure that multiple AoE masters arbitrate between each other before
putting any packets to a device. Finally, I'm pretty sure that AoE has
retransmit built in (I did look at the AoE specs after your talk, but I
may not have read it correctly).
All of that flow control is highly non-deterministic. The vendors
aren't particularly happy with it either:
http://www.networkworld.com/netresources/0913flow2.html
You are correct that a star topology switch could conceivably implement
a round-robin, token-like arbitration. But those guarantees will not
cross the Ethernet link between switches. In addition, if you are going
to do that, you might as well *actually* change to a loop configuration
and get the QoS guarantees.
-a
--
[email protected]
http://www.kernel-panic.org/cgi-bin/mailman/listinfo/kplug-list
