Recommendation for gigabit NIC
nelson18 at llnl.gov
Tue Apr 27 11:38:01 CDT 2004
On 14-Apr-2004 Peter Mueller wrote:
>> Don't get the Broadcom. We've had numerous problems with it that
>> are hardware related (i.e. not fixable via drivers)
> Can you extrapolate a little? I've got a few of these running on my
> Poweredge Servers, using tg3.
We have 16 1750's, each with the Broadcom NIC and a 2650 acting as the master.
We haven't had any problems with the NIC in the 2650. For the 1750's, we've
tried both the tg3 drivers and the bcm5700 drivers and the results were similar.
Basically, when the I/O channel reaches 2GB of traffic, things stop for awhile.
After 50 seconds, things start working again. After that point, things time out
for 50 seconds at random intervals.
I've seen numerous reports of this in the thread. Switching drivers or
resetting drivers probably only helps in that it is resetting the I/O
counter to zero.
After awhile of this nonsence, I just popped in Intel Pro/1000 MT cards
(e1000 drivers) and everything has been fine. Our Computer folks (not the ASCI
folks, but the Linux clustering folks) have noted similar things and determined
that it was a hardware problem with the Broadcom units. No info on if it is
*all* Broadcom NICs in gigabit mode, or just some of them, or power supply
issues sending too high/low voltages to the NICs, etc. etc. etc. But at this
point, I decided I couldn't waste any more time dealing with it.
I've seen threads saying that tg3 is better than bcm5700, threads saying that
bcm5700 is better than tg3, some saying they also have the 2GB issue, others
seeing the 50 second timeout, others saying that the cards don't
autonegotiate at 1 gigabit, and still others saying that *this* driver includes
thus-and-such that in the past it didn't and so is now better than *that*
So if your broadcom NIC card is working for you, that's great. You may want
to ask the question, do you know if it is really working? i.e. are you hitting
Hope this helps...
Scott D. Nelson B543 Rm2236 3-1250
Group Leader, Accelerator Design and Code Group
Lawrence Livermore National Laboratory
7000 East Ave., L-645 Livermore CA 94550
email: nelson18 at llnl.gov http://www-eng-x.llnl.gov/dsed/
More information about the Linux-PowerEdge