[vortex-bug] Asymetric throughput with 3c905b NIC

[=?iso-8859-1?Q?C=E9line_and_R=E9gis_Duchesne?=]

Hi,

I'm have 2 machines (A and B), both Pentium II 300 Mhz equipped w/ a 3c905b
(cyclone) NIC, linked together w/ a cross-over cable.

I use "ttcp" to measure the TCP throughput between them. Unfortunately, it
seems that this measure is asymetrical. I get the following results:

"->" means that the machine on the left sends the TCP payload, and that the
machine on
the right receives it.

          Machine A                    Machine B            Result (Mb/s)
          ---------                    ---------            -------------
0) vanilla kernel 2.2.14      -> RedHat kernel 2.2.12-20    OK  (93)
1) vanilla kernel 2.2.14      <- RedHat kernel 2.2.12-20    Bad (50)

2) vanilla kernel 2.4.0-test1 -> RedHat kernel 2.2.12-20    OK  (94)
3) vanilla kernel 2.4.0-test1 <- RedHat kernel 2.2.12-20    Bad (39)

4) vanilla kernel 2.4.0-test1 -> vanilla kernel 2.4.0-test1 Bad (18)
5) vanilla kernel 2.4.0-test1 <- vanilla kernel 2.4.0-test1 Bad (25)

6) vanilla kernel 2.2.12      -> vanilla kernel 2.4.0-test1 Bad (37)
7) vanilla kernel 2.2.12      <- vanilla kernel 2.4.0-test1 OK  (94)

8) vanilla kernel 2.2.12      -> RedHat kernel 2.2.12-20    OK  (93)
9) vanilla kernel 2.2.12      <- RedHat kernel 2.2.12-20    Bad (40)

This experiments are perfectly reproducible.
As you can see in 2) and 7), this is not a defect in the cross-over cable.
It seems to be only related to the driver versions I use.
Whatever the throughput is, ifconfig eth0 doesn't report any error or
collision.

A friend of mine told me that he solved a similar problem by using 3Com's
driver for this card, instead of Donald (and al.)'s one. I couldn't verify
this, because I can't "ifconfig eth0 up" with 3Com's driver (it reports
"SIOCSIFFLAGS: No such device", which sounds silly
since it does a SIOC_G_IFFLAGS right before that, with the same device
name...).

Does anybody have a clue what's going on here? I can provide vortex_probe(),
mii-diag outputs if needed, and I'm ready to test any patch you will send
me.

Please reply to hpreg@vmware.com, as I'm not subscribed to any of those
lists, and I have no idea how the hell I can teach Outlook Express how to
send a "Reply-To:" field :)

Thanks for reading, and thanks in advance for your help,
--
Régis "HPReg" Duchesne
Member of Technical Staff
http://www.vmware.com/