[72319] in North American Network Operators' Group
RE: Quick circuit question.
daemon@ATHENA.MIT.EDU (Michel Py)
Tue Jul 6 13:48:56 2004
Date: Tue, 6 Jul 2004 10:43:33 -0700
From: "Michel Py" <michel@arneill-py.sacramento.ca.us>
To: "Drew Weaver" <drew.weaver@thenap.com>, <nanog@merit.edu>
Errors-To: owner-nanog-outgoing@merit.edu
> Drew Weaver wrote:
> is duplex a consideration with lines of these types?
DS3s are full-duplex AFAIK. There are two coax cables (one in one out)
and I don't ever remember seeing anything that looked like a collision
detection mechanism.
Compare to Ethernet: on 10base-2 or 10base-5, you have half-duplex as
the physical medium (single coax) is the same for transmit and receive.
Same on a 10base-T or 100base-T hub, where even if the TX and Rx pairs
are separate they are mixed inside the hub.
When connected to a switch, a NIC that either is configured to
full-duplex or negociates full-duplex actually deactivates its built-in
collision detection circuit.
> We have 1 DS-3, that if we get close to 45Mbps IN/OUT the
> other direction will be completely unusable.
The condition you describe is rather common: your upstream is maxed out
and your downstream becomes very slow. This has nothing to do with
duplex: your downstream sits unused. The perceived slowness is likely
due to DNS requests, TCP SYNs and the like being slowed or discarded in
the egress queue.
If you traffic-shape your egress (by tweaking queuing, among other
things) you could get close to 45Mpbs in and out. Usual tricks include
granting small packets a higher priority, and lowering the priority of
packets that max out your pipe. Things such as wRED and dCAR are to be
looked at too.
Michel.
