Hi,
I'm trying to get an understanding of the flow of packets when openvswitch is
used in an offloaded lookup setting - NIC/Switch.
1a. Is it right to assume the packets still pass through the kernel datapath,
but there is no lookup? Instead it's just handed off to the device?
1b. Or is it that the OVS kernel module is not used at all? The packet is sent
to the device directly from the stack and the device has to send the first
packet back to the daemon in the hypervisor to learn and configure the switching
on the NIC?
2a. In the case of a NIC with switching on chip, I'm guessing the advantage of
offload is only realized if there is more than one port, is that correct? i.e.
if there are multiple single port devices on openvswitch, it's better to just
not do the offload since we can't send packets from NIC1 to NIC2.
2b. Offload also means the latency between VMs on the same system is higher?
3. With offloaded NICs, it's sort of clear to me that OVS will run in the
hypervisor and switching will occur in the device. But how exactly does OVS work
with switches? Does OVS run on the processor of a switch? And the first packet
of a flow is sent to the processor to learn/add rule and this is written to the
switch? In this case the kernel part of OVS is unused. Is that correct?
4. OVS and SR-IOV seem to be mutually exclusive. Since the hypervisor does not
see the packet, OVS obviously can't switch on it. Or if 1b is how it works, then
SR-IOV can also work with OVS?
Thanks,
Nithin.
_______________________________________________
discuss mailing list
[email protected]
http://openvswitch.org/mailman/listinfo/discuss
