[kwlug-disc] CAT6 - worthwhile? OT: switches, duplex, collision domains, and throughput

[Gordon Dey]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 13-11-24 10:10 AM, John Johnson wrote:
> Re: "each device on copper has a full speed dedicated connection"
> 
> AFAIK not necessarily While with point-to-point copper there is no
> "ether" as there is with wifi the hardware internals in a router or
> switch may include a common bus. i.e. shared media inside the box.

I agree that there is a common switching frame bus inside of a
"switch". Example:
> https://www.vitesse.com/products/images/VSC/big/vsc7398.png

A "copper", or UTP/RJ45 host-switch connection is a point-to-point,
collision-free, "domain", or network, because the transmitter and
receiver don't share a single circuit. Often, host and switch can even
negotiate a "full-duplex" connections, where they both transmit
simultaneously. I think this is what the first line was trying to get
at. No collisions means no time wasted in back-off algorithms or
protocols, and simultaneous transmission means no time wasted in
waiting for acknowledgement between transmits.

If all of the network path pieces A-C or B-D support jumbo-frames (
9,216B per transmit
http://www.networkworld.com/community/blog/jumbo-frames ), then the
ratio of payload to overhead bytes drops even further, leading to
further increases in network throughput. 802.11 wifi supports a
maximum 2,346B per transmit (
http://forums.wi-fiplanet.com/showthread.php?9539-802-11-frame-size ).

I suggest that if one compares the same Mbps wifi connection
throughput to a collision-free, full-duplex, wired connection, the
wired one wins. What Mbps we're comparing, I think, is the cat 3 vs. 5
vs. 5e vs. 6 UTP discussion.

> I doubt that two servers (A and B) streaming data to two
> workstations (C and D) as in A to C and B to D at the same time
> through a common router / switch would each enjoy the maximum
> throughput that would be available if only one server was
> streaming.

Switching fabric chips of the last decade *can* routinely run at a
"wire speed", of 1G packet/second. Also switches (vs. hubs) learn the
addresses of the hosts on their ports, and switch non-broadcast
packets only between participating ports. One can see this by trying
to see/sniff/wireshark/tcpdump A-C traffic off of either B or D's port
on the switch. Only traffic such as Ethernet spanning-tree broadcasts,
ICMP ARPs and IP broadcasts appear.

I agree A-C traffic may not see full bandwidth with the presence of
B-D traffic, but if the streaming traffic is mostly non-broadcast,
than they should be mostly isolated and the degradation in throughput
seen,  is hardly noticeable.

IMHO,
Gord

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.10 (GNU/Linux)
Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/

iQEcBAEBAgAGBQJSkqeWAAoJEGgksGZ7nTPG+xoH/A65zaFdjx56cVEpmHsWvMBo
kP9ZCcFZ3T/YNn1/NW/h+YBtPQvOEc9hCu8haVNBUgrlhS13Ut0lhSbrNPkyuGhU
+P6OvOoiVO61L9uu+5q77RCFHVlZxHVo0GkytDdwjvU3bvO44T3ol6bbdgTo79z0
40r+jKXELeadI3TcRu1x2Pef+9kmvoLk2K7rUDsXiEsYB+pK/8oLAEeu/Fhn9HeW
ixp4CZQtGAKo3cglqgKjN5s5mYuZmL0P0MIpI6bOuVP/j1kAJyMFEWeNN869Vzzq
+0ylYvc4kQSrS72DizjQvbwBPS+7tRhNti+LRx1UVAHPnzMw7MyVxh4UQiUH5Tc=
=muIs
-----END PGP SIGNATURE-----