Massive Packet Loss Between VMs on XCP-NG

I have a strange problem that’s driving me mad, hoping someone can share some insight.

Hardware: XCP-NG 8.2.1 on a Dell R610, dual X5675s, 96GB DDR3, RAID6 SAS array. Onboard NIC is a Broadcom NetXtreme II BCM5709. Running roughly 7 VMs with low CPU usage.

Using iPerf3 I’m observing a huge number of retransmissions between virtual machines, surely more than is normal. This is between two virtual machines on the same host, with the network tied to the (disconnected) physical Broadcom NIC.

Much the same story in reverse.

I have tried updating firmware, which if anything made it worse, and all offloading on PIFs and VIFs is turned off using this script: Blog/Disable Network Offloading on XenServer and XCP-ng - Forza's ramblings

Which seems to help marginally.

What’s even stranger is that I don’t think this problem exists / is as pronounced when traffic is traversing a PIF (although I don’t currently have any iperf numbers to support this). Before I pinned this packet loss as the issue, I only knew that with my web applications running as VMs in XCP-NG, HTTP transfers/openspeedtest topped out at 500Mbit/s. When I moved them to a separate host over a 10G link, the transfers ran along at about 1.1Gbit/s. Not great but I think I expect too much of this old hardware.

I’m also running pfSenese virtualised (with offloading disabled) which has these iperf results:

In reverse:


I can’t make head nor tails of this. If anyone has any suggestions on either a fix or where to start troubleshooting I’d love to hear it.

Cheers, have a lovely day.

The most likely issue is the “Broadcom NetXtreme” card. Broadcom cards are just terrible, we just helped a client solve a weird XCP-NG issue across 20 high end servers at a datacenter because of those cards.

1 Like