[vortex] DIFFERENT MAC ACCESS

[Markus Fraedrich]

> I only read once sometime ago some docs on the Scyld site which detailed
> how data is put on the wire, so I'm in no way expert in this. I believe
> that if you use a switch, the IFG only affects the physical link
> between the card and the switch, but will in no way be related to other
> links (between the same switch and other cards). I have no ideea about
> what happens when you use a hub.
> Not related to IFG: I think that there are switches were you can define
> QoS for certain ports and/or types of traffic. This might in fact work
> much better for you.
>
A switch has the advantage of making direct connections between two link
partnes (full-duplex access),
then there is no csma/cd access and the ifg doesn´t matter. But with multi-
or broadcast connections,
I´ve got the same collisions problem. 3com switches include the PACE
technology with a modified backoff algorithmus (defined by the port -dont
really know) for making it more effective for video streams.
 But I want to realize a low cost implementation, with changing the MAC
access (
setting the IFG much higher) on the NICs.
It seems to be working quit good with a hub.

> So you want to change IFG in real-time depending on whether the packet is
> part of a stream or not ? This sounds very bad - drivers are actually
> trying to get away from using I/O within the hot-path and you want to
> introduce (quite a lot) more I/O ?
>
I know it sounds bad. At the moment I´m setting the IFG with a simple IOCTL!
And i´m really searching for a solution to set this iocl comand depending on
the packets to send.



> I'm not familiar with the in-kernel QoS. However, it's done at higher
> levels than those that we discuss here - you will not be able to skip
> packets the are currently in the driver Tx queue, for example. Maybe
> you should consider Don's suggestion: replace the card with a faster one
> (Gigabit) - then even if you do QoS at higher levels, at least the driver
> Tx queue will get cleared faster.
>
Gigabit it so expensive.
The Kernel QoS has the advantage to sort packets before they are transmitted
into the NIC,
perhaps it is easier to do this, instead of making these with the driver.



> Another suggestion: a hard real-time OS.
>
I thought of this also.

Markus Fraedrich
> --
> Bogdan Costescu
>
> IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen
> Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY
> Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
> E-mail: Bogdan.Costescu@IWR.Uni-Heidelberg.De
>
> _______________________________________________
> vortex mailing list
> vortex@scyld.com
> http://www.scyld.com/mailman/listinfo/vortex
>