[eepro100] Intel HPQ and MAC Setting question

[Donald Becker]

On Tue, 27 Nov 2001, Markus Fraedrich wrote:

> I did some tests with Realtek,3Com and  SUN NICS and it seems that setting a
> higher IFG (more than a slottime)
> could give one client a deterministic bus request.

Nope.  You have only changed the definition of "fair" in "fair access".

You must understand Ethernet, which is surprisingly complex and robust
for such a simple protocol.

Image a new transmit request being generated at a point that causes a
collision with the (single!) high priority station .  Your high priority
station will back off, and then effectively have a larger IFG than low
priority stations.

So while the network will accept more average traffic from the normal
IFG machine, you still won't have deterministic access.

[[ In years past someone would have responded with "use Token Ring".
The correct response is that token ring doesn't have deterministic
access time either.  When the token is lost, as frequently occurs, the
access time can be really bad.  The probabilistic access time graph is
just different, in a way that is worse for most application. ]]

> This QoS is for a small home network with about max. 5PC and a connection
> with a normal hub, the Qos
> is for realizing a DVB stream over  Ethernet.
> Know i want to implement a high priotization queue for video stream.Therefor
> I try to understand the building
> and controlling of the buffers.

To do this reliably you will essentially reproduce the complexity of an
ATM switch controller.  ATM was supposed to be a simpler system than
Ethernet, but it's now accepted as a much more complex due mostly to the
_required_ QoS handling.

You can optionally achieve the same effect with Ethernet by using
switches, packet scheduling, and per-link bandwidth allocation.

The cost/value relationship of precise QoS, vs. "throw more cheap
bandwidth at the problem" is demonstrated by how many QoS Ethernet
systems you find in the market.  The preferred solution is link
flow-control, which is now silently implemented even on the least
expensive NICs and switches.

Donald Becker				becker@scyld.com
Scyld Computing Corporation		http://www.scyld.com
410 Severn Ave. Suite 210		Second Generation Beowulf Clusters
Annapolis MD 21403			410-990-9993