[e2e] packet loss in the Internet

[John Kristoff]

On Thu, May 23, 2002 at 08:22:44PM -0700, Stefan Savage wrote:
> I agree about congestion.  However, I've measured significant loss in
> enterprise networks on several occasions that arose from Ethernet
> full-duplex/half-duplex conflicts (its still not clear to me if the
> Ethernet autonegotiation protocol is fundamentally inadequate or if
> current implementations are simply poor).  Its unclear to me how
> prevalent this problem is, but I've run into it in three separate
> networks.  

It doesn't seem as prevalent as it used to be in my experience.  The
most common scenario is when one end of the link is manually configured
for speed and duplex, while the other is set to auto-negotiate.  Since
manually setting the link parameters disables A-N, the other side will
be unable to detect the proper parameters and often fail.  One end will
often default to 100 Mb/s half-duplex, which is rarely correct when the
other side is manually set (typically when manually set to 100 Mb/s full
duplex I believe).

>From what I understand, the spec does require some specific timing
requirements for the process of A-N.  Current hardware I've seen
seems to indicate that vendors are implementing it properly, but
that may not always have be the case.

A couple other notes.  Auto-negotiation does not take into account
the quality of the link.  So if the link media is poor, they may
negotiate to 100 full duplex, but may actually perform poorly due
to media imposed errors.  Also,  A-N will only be performed during
during NIC initialization (or re-initialization).  So if someone
happens to change say a switch port from one to the other, the
other end will probably need to be manually changed or reset also.

Static devices and links are generally best set manually to the highest
compatible speed and duplex setting.  This might include switch to
switch, switch to router or switch to server connection.  Others such
as switch to end station PCs are generally OK with auto-negotiate
enabled on the switch if the PCs have fairly currently hardware and
NIC drivers.

To answer your question... current implementations are probably OK,
but I think you can easily argue that the protocol is fundamentally
flawed.  Not from a standards or engineer's perspective, but from the
end user or network administration perspective.  Is there an e2e
lesson here?

John