Re: NIC on morollan
Another data point:
We just recently ran into the problem with the e1000 and it was only
when we were doing more than running a beacon client. In our case,
we showed zero loss for our beacon clients but when we ran a beacon
server, the server (on a machine with an e1000) showed 40% loss of
received multicast from the other sites. THis pattern repeated for
all e1000 and e100 network interfaced machines we tried. If they ran as
a client (small closed network of 6 clients), then they showed zero
loss. If they ran as a beacon server then they showed a 25-45% loss
of reception of all of the other clients' multicast traffic.
When we ran the server on a machine with another vendor's 10/100
network card, there was no loss at the server. We are now in the
midst of trying a different vendor's gigE card to see if one of the
original machines that exhibited the problem is fixed by using a
different network interface.
Deb
Derek Piper wrote:
>
> Hi,
>
> I installed the beacon on my workstation
> (dcpiper.informatics.indiana.edu) and have not noticed any difference
> running for a number of hours yesterday (with checksum offloading
> disabled) and today (with checksum offloading enabled-default). I get no
> loss whatsoever. I will talk about this with our network people since
> the two machines are within the same room, though not necessarily the
> same switch. I'll find out more and let you know of any 'revelations' :)
>
> Derek
>
> PS: The beacon page has an awful lot of 'NA's all over it. Someone kill
> the US-Australia link?
>
> David Gwynne wrote:
>
>> From: "Mitch Kutzko" <mitch@ncsa.uiuc.edu>
>>
>>> At 01:52 AM 10/27/2004 +1000, Christoph Willing wrote:
>>>
>>>> Mitch,
>>>>
>>>> I see that you're the contact for morollan.ncsa.uiuc.edu - can you
>>>> confirm/deny whether its NIC uses the Intel e1000 chip? Its output in
>>>> the central_loss page shows a pattern of losses which we think can be
>>>> attributed to this driver/chipset.
>>>>
>>>> chris
>>>
>>>
>>> Hi, Chris -- Sorry to provide you with a "bad" data point for your
>>> theory,
>>> but here's the info for the NIC on morollan:
>>
>>
>>
>> Perhaps this isn't a bad data point after all. Both the intel gigabit and
>> the 3com 3c905 tornado chipsets have support for tcp/udp checksum
>> offloading. In theory though, this should help improve responsiveness
>> since
>> the host computer has to do less work processing the packets. I
>> believe the
>> linux drivers support this offloading for both controllers. I know it
>> is on
>> by default in the e1000 driver.
>>
>> I have tried disabling the checksum offloading with the e1000 driver
>> which
>> really did improve the pattern of loss visible on the central_loss
>> page, but
>> didn't eliminate it entirely. I would have liked to have attributed
>> the loss
>> to network conditions, but other hosts on the same subnet using a
>> broadcom
>> gigabit chipset experienced no loss.
>>
>> David Gwynne
>>
>>
>>> [root@morollan root]# cat /etc/sysconfig/hwconf
>>> [....]
>>> class: NETWORK
>>> bus: PCI
>>> detached: 0
>>> device: eth0
>>> driver: 3c59x
>>> desc: "3Com Corporation|3c905C-TX/TX-M [Tornado]"
>>> network.hwaddr: 00:04:75:AB:2A:5C
>>> vendorId: 10b7
>>> deviceId: 9200
>>> subVendorId: 10b7
>>> subDeviceId: 1000
>>> pciType: 1
>>> pcidom: 0
>>> pcibus: 2
>>> pcidev: c
>>> pcifn: 0
>>> [....]
>>>
>>
>>
>>
>>
>
--
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Deb Agarwal e-mail:DAAgarwal@lbl.gov
MS50B-2239 phone :(510)486-7078
Lawrence Berkeley National Lab URL: http://www-itg.lbl.gov/~deba
Berkeley, CA 94720
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~