gigabit ethernet card problems - smc9462TX

Velocet math at velocet.ca
Wed Mar 27 17:44:12 EST 2002


On Wed, Mar 27, 2002 at 03:26:58PM -0600, Alexis Zubrow's all...
> Hi.
> 
> I'm a relative newbie to both linux and beowulf clusters.  I am starting
> up a small beowulf cluster.  Until we get some more funding, I have two
> dual processor computers connected with a crossover cable.  I'm using the
> ns83820 driver in the 2.4.17 kernel.  I have not been able to get any real
> connection between the two cards.  Any suggestions would be greatly
> appreciated.
> 
> Hardware:
> Dual AMD MP 1800+, Tyan S2460 mother board, SMC9462TX ethernet cards
> 
> Tests:
> -crossover cable is good

dont use crossover. use a regular cable. GBE copper doesnt have a crossover
standard. Make sure you wire up all wires to the correct ethernet standard.
This includes proper colour matching to one of the two accepted wiring
colour standards as then you get the proper pairs twisted the proper number
of times per distance, etc.

> -if connect (with crossover cable) one of the cards to another
> (non-gigabit card), it recognizes and correctly auto negotiates to the new
> media speed.
> -ping between two gigabit cards has 100% packet loss.

There are also problems with the designs of some cards and some motherboards -
64 bit cards in 32 bit slots and vice versa on some boards DO NOT WORK at all.
For eg, the 32 bit Addtron cards I got (6 different cards) dont work at ALL in
ANY slot on Tyan Tiger 2460 boards, but work fine in any Tyan Tiger 2466
slots (tho I would heavily suggest not using the addtrons, they're quite alot
slower than the ARKs we used in our tests).

I had alot of problems with the DLINK card that Ben LaHaise (ns83820 driver
author) lent me for some testing, and a couple problems with the 2 SMCs he
also lent me. The DLINK established link with one other card once and had
errors, in about 10-20 tries over 2 months. The SMCs were abit finicky, not
working with up to half the other cards I tested. 

Despite the troubles with the DLINK I also actually went and bought a pair
each of Ark 32 and 64 bit cards for a cluster Im building now as part of a
test - they're exactly the same design as the DLink cards down to the
component model numbers on the board, and they work great in both Tyan Tiger
2460 and 2466, and talk to eachother (and to the PureData and SMC cards
on loan) with no problem, provided there's good cables in use.

Up to 10-20% of all cables we made by hand with Cat5e plenum (and non plenum
:) wiring (solid not stranded wire) would not give a link when plugged in, and
a further 5% did not perform optimally despite link light (either no pings
went through, or there were many errors in xmission). Test your cables
heavily, GBE is extremely sensitive - test even Cat6 cables which are
supposedly rated for this - we tried a bunch (higher priced) and about 5-10%
had troubles of some sort with our setup (... despite their higher price :).

Your problem may well be the cards may not work in that board like my addtrons
(though I did use SMC cards in the 2460 succesfully, not sure if they were the
same card model # as yours). The ark 32 bit cards are $43ish USD at last check
from comready.com (plug: which I ordered much equipment from that was not
available in canada, with absolutely no problems and excellent service - even
allowing me to return the garbage addtron cards). The 64 bit cards were
about $66-68 or so.

/kc

> 
> 
> thanks for your help.
> Alexis
> 
> p.s.  I know this may be a bit too hardware specific for this group, but I
> thought that this group would have the most experience with gigabit cards.
> 
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf

-- 
Ken Chase, math at velocet.ca  *  Velocet Communications Inc.  *  Toronto, CANADA 
_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list