After reading some posts and blogs on vSphere5 and E1000E performance my curiosity was triggered to see if actually all these claims make sense and how vSphere actually behaves when testing.
The setup I used is similar as described in http://www.vmware.com/pdf/vsp_4_vmxnet3_perf.pdf. The setup looks like:
Bare metal server (Client): B22-M3, 16GB, 2xE5-2450, 1280VIC
vSphere ESXi 5.0 server: B200-M3, 64GB, 2xE5-2680, 1280VIC
To accommodate the tests, the 1280VIC’s are connected to 2108 IOM’s and we are only using Fabric A / 6248-A.
The VM is configured in the following way (screenshot):
- Local Area Connection: E1000
- Local Area Connection 2: VMXNET3
- Local Area Connection 3: E1000E
- 4GB Memory, 1 vCPU
- Windows 2008R2
The following is a result of the best performance test I did run
Average transmission in Mbit/s on Windows
Average CPU load on Windows measured
%USED counter in esxtop
MbTX/s in esxtop
Tx Bit Rate in Gbps as seen by the 2108 IOM module
|Adapter||Win Net||Win CPU||VM CPU||VM Net||FEX Net||Graph|
|Win Net||Average transmission in Mbit/s on Windows|
|Win CPU||Average CPU load on Windows measured|
|VM CPU||%USED counter in esxtop|
|VM Net||MbTX/s in esxtop|
|FEX Net||Tx Bit Rate in Gbps as seen by the 2108 IOM module|
We can clearly see that all adapters can be filled, full line speed. There are small differences but these could very much be due to sampling periods etc…
There is a higher CPU usage seen for E1000 and E1000E adapters, for both WIN CPU and VM CPU. I think however only for E1000 there is a high penalty where for E1000E this stays within acceptable limits.
I’m not a bench guy neither is this my job, hence these figures are just my personal observation and by no means are they a result of a full professional benchmark. They are however fully reproducible.
The attached graphs do show some dips, I did not further look into them. I know technically why they are there, but did not look into fixing them.