Nic Card

[Ken897]

I have a HP NC360T 2 ports NIC card, it is using the Intel 82571EB chipset. Could anyone explain to me why when the network cable is unplugged from the NIC Port, the CPU Utilization will go up to 60 to 70% for 3 to 4 seconds, causing my keyboard, mouse and application to hang for that few seconds. The OS is Windows 2003 SP1 Standard Server.

My server also come with the embedded NIC, which is the NC373i, when I unplug the nic card from the NC373i, there is no CPU utilization, my keyboard, mouse and application have no issue.

At first I thought maybe due to the Intel and Broadcom driver.

But my other server has the NC7170, which is using Intel driver, but when I unplug that nic cable from the NC7170, there is no CPU utilization neither. But the NC7170 is a PCI-X adapter.

I only have one PCI-Express NIC to try. Not sure is there anyone have a PCI-Express NIC using Intel or Broadcom driver. When unplug the NIC cable, the CPU utilization will go up to 70% for few seconds or not? If yes, why it will generate so high CPU utilization, while the other nic controller has no such behavior.

[nmX.Memnoch]

Just a question...what's the reason for removing the cable so often?

I have an Intel PCIe NIC but I've never noticed that behavior since...well...I really don't have any reason to remove the network cable unless I'm moving the PC, in which case it's shutdown for that.

[Ken897]

The reason to unplug the nic cable is trying to simulate nic cable fail, especially i have the two ports team up as Network Fault Tolerance, then i noticed whenever I unplug the nic cable, the CPU utilization went up. It happened only when the adapter speed is set to 100MB/Full duplex. Cause my nic cable is connected to the CISCO 2950. Since it is set to run at 100MB/Full duplex, that's why network admin told us to set the adapter speed to 100B/Full duplex as well.

I also team up the broadcom embedded nic as Network Fault Tolerance. It is also set at 100MB/Full duplex, but when i unplug the nic cable from the embedded nic port, it doesn't have this CPU utilization at all.

I just want to know what is the reason behind this Intel card will generate this CPU utilization, when the nic cable is connected?

[nitroshift]

Protocol handshake broken?

[nmX.Memnoch]

i have the two ports team up as Network Fault Tolerance

That's probably the answer. I would imagine it has something to do with the Intel driver negotiating the failover. I can't explain why it doesn't happen with the Broadcom NIC...other than it's non-native to the Intel fault tolerance drivers.

The Intel NICs do a very good job of offloading TCP/IP transmit and receive functions from the CPU. When the failover happens the CPU probably has to take over for a few seconds until the other Intel NIC can fully initialize. Again, this may explain why it doesn't happen with the Broadcom NIC. Broadcom NICs do some offloading as well...but again, that's non-native to the Intel drivers.

[Ken897]

Thanks for your help!! Just want to double check, does your NIC card also have such behavior? When you set the adapter speed at 100MB/Full duplex, then unplug the nic cable and the cpu utilization will go out for few seconds. But when we set the adapter speed back to automatic, then unplug the nic cable will not have any cpu utilization.

[nmX.Memnoch]

I really don't have the ability to test that for your specific setup since I only have the one NIC in my machine. I suspect it may very well though.

Which Intel driver version are you using? They've been releasing new ones like crazy lately.

What really gets me is all these net admins forcing switches to 100/Full instead of just leaving them on Auto. It's like a left over from YEARS ago when Auto didn't exactly work like it was supposed to.

I hope they fully understand that the GigE standards calls for Auto or nothing to run Gig speeds. I'm fighting this battle with our config management shop at work right now as well. They have some ports left on Auto but most are on 100/Full. We have a policy of forcing the NICs to 100/Full when deploying a workstation...and then waiting to see if the user complains about slow network before we know if the port is actually on Auto or not. What really ticks me off is that when we finally deploy GigE (within the next 6 months I hope) that we'll have to go around and manually set all of those NICs back to Auto.

Heh...we even have one of our net admins trying to tell me that 100/Half is faster than 100/Full!!