Under some circumstances, the EtherExpress Pro 100 can destroy networks. Discussions of this on the mailing list hosted by GSFC show that many folks have had different variations on this. In my case, an Intel motherboard, L440GX, which contains an 82555 Ethernet chip, will spew massive numbers of duplicate packets when using multicast. This is consistent with some software but not all, which tickles the multicast driver functions. It's not yet clear why. An interesting note is that enabling promiscuous mode (either by ifconfig eth0 promisc or by running tcpdump) will stop the flood of duplicate packets. The duplicate packets appear to be the result of collision problems (i.e. the chip apparently thinks it has a collision, and resends the packet lots of times. In my network, anything on the same segment as this one server will croak from overload. Anything on other Ethernet switch ports will see only the expected traffic. In my case, the server was running at 100Mbps, half duplex. This is a complicated set of issues, and one which appears to be ongoing, based on the info on the GSFC Beowulf website. Since Intel claims this particular motherboard is certified for RedHat 6.0, perhaps they'd be willing to help debug the drivers for this card?
Assigned to dledford
Current versions of the eepro100 driver are thought to work much better. There is one thing that can cause the type of behavior noted beyond just a broken driver, it can be caused when the driver thinks the card should be using half duplex but the switch the card is plugged into is using full duplex and vice versa. Care needs to be taken to make sure there isn't a duplex mis-match. If you verify that there isn't a duplex mis-match and that you are using the current eepro100 driver and you still have a problem, then please re-open the bug.