Paul H. Anderson writes:
> 
> Is anyone else running two 164LX 533MHz (or greater) boxes on the same hub
> or switch both running netbsd 1.4, both with dec 21140, 21143 or intel
> etherexpress cards?  If so, do you get equal performance doing a large get
> vs a put via ftp between the two machines?

I wish I had this setup.. :)  Have you tried to use ttcp for this testing?
Something like:

box1# ttcp -r -s
box2# ttcp -t -s box1
(and the other way)
Also:
box1# ttcp -r -s &
box2# ttcp -r -s &
box1# ttcp -t -s box2
box2# ttcp -t -s box1
(last two at the same time, or as close as possible)
You might prepend 'sleep 2;' on the box1 command if it takes a second
or so to switch to the other box and hit enter.


> We do not, and for the life of us, we are completely unable to get this
> working properly.  It is remotely possible that we're doing something
> wrong, or truly are just seeing the effects of a poor driver
> implementation (e.g. the reported problems with 21143 tulip).  I know
> we're not stupid, and I know we don't have subtle hardware problems -

From a NetBSD/i386 (Celeron) to a NetBSD/i386 (PPro200) I get great
full duplex performance.
I have these two cards:
"box1"  -- NetGear FA310TX early rev, with DEC chip.
de0 at pci0 dev 16 function 0
de0: interrupting at irq 9
de0: 21140A [10-100Mb/s] pass 2.2
de0: address 00:40:05:40:f7:60
"box2"  -- original SMC EtherPower 10/100 PCI (with RJ45 + DB9 connectors)
de0 at pci0 dev 19 function 0
de0: interrupting at irq 9
de0: SMC 9332DST 21140 [10-100Mb/s] pass 1.1
de0: address 00:00:c0:35:6e:bf
de0: enabling 100baseTX port

Running "full duplex" ttcp as outlined above I see:

box1->box2:
ttcp-t: 134217728 bytes in 14.66 real seconds = 8939.49 KB/sec +++
ttcp-t: 16384 I/O calls, msec/call = 0.92, calls/sec = 1117.44
ttcp-t: 0.0user 2.9sys 0:14real 20% 0i+0d 0maxrss 0+2pf 35636+1015csw

box2->box1:
ttcp-t: 134217728 bytes in 18.56 real seconds = 7063.04 KB/sec +++
ttcp-t: 16384 I/O calls, msec/call = 1.16, calls/sec = 882.88
ttcp-t: 0.0user 11.6sys 0:18real 62% 0i+0d 0maxrss 0+2pf 3458+8402csw


> we've got 10 164LX rackmount boxes that we are trying to use as a
> supercompter cluster - they all show the same behavior very consistently.  
> This makes our cluster worth approximately nothing to us, and it is
> getting extremely frustrating.

Bummer, nice hardware though. :-)  The PCI DMA bug mentioned might be
worth looking into.

> Again, any suggestions what we might try would be greatly welcome.  If it
> comes down to it, I'll ship some of these mothers somewhere if that will
> help (or work with someone remotely).  Their next stop is pretty much the
> dumpster.

Uhmm, #define dumpster ( Andrew's Truck << From SouthWest Michigan)
:=)

Have you tried any (ahem) other OSes?  If would be interesting to know
if this is a NetBSD issue. (e.g. other OSes workaround a hardware bug?)

-Andrew
-- 
-----------------------------------------------------------------
Andrew Gillham                            | This space left blank
gillham@whirlpool.com                     | inadvertently.
I speak for myself, not for my employer.  | Contact the publisher.