High Speed Beowulf Cluster using gigabit ethernet

Donald Becker becker at scyld.com
Tue Feb 5 11:46:10 PST 2002

On Tue, 5 Feb 2002, W Bauske wrote:
> Trent Piepho wrote:
> > Addtron claims their 320T and 620T (64bit) cards work with Linux.  I don't
> > know what the ark and dlink cards' status is.  The quality of the driver (
> > stability, hardware TCP/IP checksum, jumbo frames, etc.) and any motherboards
> > incompatibilities are also a big question.

There are two independent dp83820 drivers, and varients of each driver.
My driver is named ns820.c

We don't currently recommend '820 cards.  We have seen lost packets
during tests, independent of what driver was used.  This was with
hardware produced at least five months ago so the problem may have been

> No systematic testing though. I suspect
> the driver is assuming a 66Mhz 64bit bus and messes up the card timing
> but have no proof. The card just does nothing. Doesn't hang the system
> though so it's not real painful to mess with.

An issue that we have seen is related to automatic bus sizing.

The '820 (65 bit) and '821 (32 bit) chips have the same PCI ID, and same
PCI chip rev. number.  The only difference seems to be in the
documentation and the printing on the chip.  The 64 bit chip pinout uses
all pins.  The 32 bit chip labels what would be the upper 32 bit pins as
"N/C" (no connect).

The problem is that the bit to enable 64 bit data transfers is loaded
from the EEPROM.  If this bit is programmed wrong from the factory, a 32
bit card (remember, it's the same chip as a 64 bit card) will try to do
64 bit transfers and vice versa.  Apparently some vendors, at least for
early cards, used the example EEPROM programming unmodified.  When the
default was incorrect for their design, they compensated for the mistake
(perhaps not even realizing it was a mistake) by modifying the driver to
set the bus size configuration based on the PCI subsystem ID.  But that
means only their cumtomized driver works.  Doh!

Probably more than you wanted to know, but it's a tidbit that may
someday be useful.

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

More information about the Beowulf mailing list