Working with an HP Customer who has 2 x 680c G7 blades in an HP c7000 with 2 x Flex-10 modules.
I have seen non vSphere hosts at other customer sites break 9 Gbps in similar configuration (460c G7, same network chipset)
However in this environment with vSphere 5.1, they cannot seem to break approx 5.6 Gbps from VM to VM when also going Host to Host.
If I vMotion the target VM to the same host as the source, which then only hits the vSwitch (vDS in this case) and never touches the physical NIC, They can get 8 Gbps Windows to Windows (NTttcp) or 13 GBps Linux to Linux (Netperf).
Checked the Netqueues option which is suported by Emulex and its already Enabled.
Tried enabling 9K Jumbos on vSwitch and both Windows VMs (Virtual Connect auto detects jumbo frames) and used a Virtual Connect network that has no uplinks to guarantee traffic never leaves the encloure.
None of these helped and we still can't break 6Gbps.
1) Any ideas on what else I can test or tweak to get the speed up?
2) Can anyone else out there with a similar setup (Flex-10 or FlexFabric and Emulex powered G7 or even Gen8 blades) tell me what kind of max throughput they are seeing?
Unfortunately they don't have any spare Blades at this site so I cannot setup 2 Physical boxes without vSphere in order to narrow the scope.
NTttcp was run with 4, 8 and 12 threads all bound to a single CPU and Receiver VM bound to a single CPU as well.
Netperf was just run with the -h switch to specify the hostname. I beleive anything past that was left default.
-Dan
C:\> ntttcps -m 12,0,ip.of.target.vm -a 4 -l 256000 -n 30000
...
Total Throughput(Mbit/s) = 5616.515
