I have no idea if this is the source of your problem (I wasn't using bonded interfaces), but it's sufficiently similar that you might want to try it.
I had a lot of problems with the network stack on VMs, both under VMWare ESXi and Xen where the network would just go numb. After a lot of splunking I determined that it seemed to be related to faulty TCP segment offload. Generally speaking, between the VM, the virtual NICs, the hypervisor/host, and the physical network card, some levels figured that they'd offload segmentation handling to a lower layer, the lower layer wasn't doing it, and the upper layer thought that it was.
Under low network load everything seemed fine but as the network got pushed things would blow up and go numb.
Turning off TSO in the VM seemed to do the trick, although I think in the Xen case I turned it off in the host as well.
The basic command is: /sbin/ethtool -K ethX tso off
While I had the above command in rc.local, I would also run the attached script in /etc/cron.hourly as there were some circumstances where tso would get reenabled.
Good luck
Devin