[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Getting Started

> I have heard, although I cannot speak from experience, that the new
> GigE-over-copper cards deliver less than 600Mbps in practice under
> Linux.

Well, I suppose the 100 Mbps cards don't give you nominal band either.
Anyway, prices are coming down all the time so, if you want to develop now
the use of a technology which is likely to be run-of-the-mill stuff in a
few years, Gbit may still be the way to go...

> If that is the case, Prof. Itti claims that you get arithmetically
> better performance with progressive addition of Realtek (100Mbps) cards
> with channel-bonding up to a point.

I think that the Realtek cards are quite bad, though inexpensive. Is there
any evidence that it is better price/performance wise to use these as
opposed to one of the really good cards like 3COM 3C905, Digital DE500 or
its many Tulip clones, or Intel EtherExpressPro?

> That point is determined more by shared interrupt problems than Data
> Link Layer issues. This begs the question; how many cards can you
> channel bond if you had interrupts to spare?

What about this ability the kernel has to deal with several PCI cards on
the _same_ interrupt line? I suppose that several PCI network cards on a
single interrupt line and a bonding driver should work more or less like a
single faster card?

> Heat removal has certainly become a cottage industry. I have seen many
> overclocker sites that have various suggestions (other than good heat
> sinks / fans / thermal compound) on ways to get your CPU core down a
> degree or 2 more by leaving the case side off or some other inexpensive
> tricks.

Well, for those out there using K7's and having trouble, overheating isn't
the only problem, watch the voltages of your power supplies too. Recently
we had two of ours drop the CPU core voltage by 10% (from 3.3V to 3.0V),
which caused hardware hangs after a period of operation ranging from a few
minutes to a few hours.

> Well, this is the issue of the day, IMHO. Every subcomponent of the
> Beowulf archetype has been improving at the pace of Moore's Law except
> for the high-speed interconnect. I would certainly prefer to use a
> protocol at the Network layer because of the application programming
> implications but if there was something that would give us lower latency
> for the same bandwidth it would certainly be a hit...

Well, there is a group in the University Computer Center here which is
going to try a 30-node machine with Gbit and the bypass of the kernel TCP
stack. Their idea is to have two network cards on each node, a 100 Mbps
one for normal TCP networking and a Gbit one for inter-node communication.
Remote boot it to be done using the 100 Mbps cards, since there still
aren't any Etherboot drivers for Gbit. They have chosen Gbit as opposed to
Myrinet, I'm told. I know several people on the team and, if this really
works out, I will let the list know...

        Jorge L. deLyra,  Associate Professor of Physics
            The University of Sao Paulo,  IFUSP-DFMA
       For more information: finger delyra@latt.if.usp.br

Reply to: