Since it's a production machine, I had to try a remedy first: I changed one network card (pub, of course) to e1000 and now it's up from 62 hours (maybe a record!)
Here is the output for the other card (virtio): I guess parameters should have been the same for the affected interface too (they are now the same for the tap interface connected to e1000 driver, and reflect the underlying physical interface). Offload parameters for tap-solariconsi: rx-checksumming: off tx-checksumming: off scatter-gather: off tcp-segmentation-offload: off udp-fragmentation-offload: off generic-segmentation-offload: off generic-receive-offload: on large-receive-offload: off I cannot switch back the virtual nic just now. I will try to increase debug on other machines, but most are 64bit Win2k3 OSs, so the virtio-net driver is not the same (it's a 64bit version at least...). Soon or later I will have some debug data to report to the list. In the meantime if you have some configuration advises, feel free to post them. Thanks, Mario 2013/11/24 Yan Vugenfirer <yvuge...@redhat.com> > Hi Mario, > > Can you check the offload settings of the tap device that is connected to > guest? > > Run “ethtool -k tap-solaripub”. > > On the guest. Raise the log verbosity by going to device manager -> NetKVM > device -> Advanced tab -> Logging.Level and changing it to 4. Use DebugView > to record the driver tracing (enable kernel trace): > http://technet.microsoft.com/en-us/sysinternals/bb896647.aspx > > > Best regards, > Yan. > > > >