Linux-PowerEdge digest, Vol 1 #741 - 10 msgs Steve_Boley at
Thu Dec 5 17:30:00 CST 2002

Can you package your fixes into a source rpm that can be installed and then
someone can run rpm -bb and build rpm or rpm -bi and then it overwrite and
compile the newer tg3 and then install it without having to get a whole new
kernel or otherwise recompile it?  Much more friendly option for people to
use and accept than a wholely test kernel that hasn't been officially been

-----Original Message-----
From: Jeff Garzik [mailto:jgarzik at]
Sent: Thursday, December 05, 2002 4:57 PM
To: Eric Rostetter
Cc: Jeff Garzik; linux-poweredge at;
bschelst at
Subject: Re: Linux-PowerEdge digest, Vol 1 #741 - 10 msgs

On Wed, Dec 04, 2002 at 01:08:43PM -0600, Eric Rostetter wrote:
> Quoting Jeff Garzik <jgarzik at>:
> > Two comments:
> > 
> > 1) Never force 10/100 full duplex on the NIC, it is _usually_ wrong.
> > 
> >
> >
> Two year old quotes about workstations (when my line is from a server)
> knowledge of my setup don't really help...

Double-red-herring here.

These are facts that exist due to ethernet, and do not change due to
age.  And this obviously impacts servers, particuarly servers with any
amount of network load at all -- precisely the case when ethernet flow
control becomes a crucial factor.

This applies intimately to servers.  Today.

> At the end of the month when we throw away all our switches and replace
> with Cisco 6xxx/4xxxx switches, I'll let it autonegotiate.  Until then,
> have to realize that old switches don't always autonegotiate correctly
> example, with cisco 55xx switches it is often a matter of which boots
first --
> the server or the switch -- as to if the autonegotiate works or fails).
> forcing settings can be a "good thing" in some cases.

Note that Cisco has _finally_ changed their docs to indicate that
forcing settings is not a good thing.

However, it is also noted that (a) old Cisco docs exist :) without this
info and (b) old Cisco switches sometimes simply do the wrong thing when
it comes to autonegotiation.

It is crucial to examine the effects of lack of flow control when you
have a bunch of NICs on the network competing for the same resources.

> > 2) I wonder why tg3 driver isn't being used?  Folks normally report
> > higher performance with tg3 versus bcm5700.
> Yes, it performs faster for about the first 1-3 hours until the machine
> hangs solid, at which time the performance goes to zero.  When RedHat
> releases a kernel with a working tg3 driver, I'll switch to it.

this issue has been fixed.  You can download the rawhide kernel with the
updated driver.

> Please don't recommend using the "beta" or "unsupported" modules on a
> production machine!

I'm not.  Please don't be so presumptuous.

> > Our bug reporters are
> > showing that you simply cannot get wire speed with bcm5700 in many
> > situations, where tg3 does do wire speed.
> Not with the latest RH 2.4.18 kernel, where it does zero speed (after a
> short while).

You apparently do not have the latest kernel, see above.


Linux-PowerEdge mailing list
Linux-PowerEdge at
Please read the FAQ at or search the list
archives at

More information about the Linux-PowerEdge mailing list