On Sat, 19 Jul 2008 21:32:42 -0400, Dan Halbert wrote
You mentioned these are Supermicro X7DBN boards. They use the Intel
(ESB2/Gilgal) 82563EB Dual-Port Gigabit Ethernet Controller. There's an open bug here: https://bugzilla.redhat.com/show_bug.cgi?id=403121, "e1000: issues with Intel ESB2/Gilgal (82563EB)". It doesn't describe your problem, but complains about other issues with that NIC, and references related bugs.
Yes, I looked at the buglist for the driver and didn't see anything related. The NICs actually work just fine at moving data. And I have the same NICs on several other Supermicros that do not have this problem.
Just for fun, I ran a backup on one of the machines, and not only did the Ethernet work well, but the "phantom" load went away while a real load was running. That's what leads me to suspect a kernel or timer bug of some sort. There was a post on the Linode site about a year ago about something that smells similar: http://www.linode.com/forums/archive/o_t/t_2729/strange_load_average.html
But those guys are doing virtualization and using newer kernels that what CentOS is distributing.
I wonder if anyone else has seen this problem?
Thanks, --Bill