single threaded test reporting poor performance with E5-2670 in VM

[ckbrou]

I have a VM with 2 vCPUs running on ESXi 5. There are two hosts. One with older X5460 procs and the other with E5-2670 procs. The single threaded CPU benchmark surprisingly scores almost twice as high when the VM is running on the host with the X5460 procs.

If I change the VM config to just have 1 vCPU and re-run the tests on each host I get the same result on the X5460 host but double the score on the E5-2670.

Why would the single threaded test be affected negatively with the VM having 2 vCPUs while running on the E5-2670 host but not when the VM is running on the X5460 host?

[David (PassMark)]

Would be interesting the run the single threaded test directly on the host machine to see the scores. I assume there is no other load on the host machines or in the VMs? Having the actual benchmark numbers from the VMs would have also been good.

I suspect it might be something VMware is better placed to answer, but here is some speculation.

The E5-2670 has hyperthreading, turbo & better power saving features (underclocking). The X5460 doesn't have most of this.

So maybe the E5-2670 never really ramps up to full turbo speed. With 2 vCPUs and only 1 thread running the average system load will be at most 50%. Or just 6% CPU load from the point of view of the physical host. With just 1 CPU available the CPU load should be ~100%. So maybe in this case the clock speed ramps up.

Or maybe the 2 vCPUs you get are really 1 physical core with hyperthreading. A hyperthreaded core isn't nearly as good as a real core. But I don't think this would explain a 2x difference.

This is just a guess however. You would need to investigate deeper to really know what is going on.

[ckbrou]

Here are the single threaded results:

VM w 1CPU X5460: 1240
VM w 2 CPUs X5460: 1208
VM w 1 CPU E5-2670: 661
VM w 2 CPUs E5-2670: 1308

These tests were done with a quiet VM on a host with no other VMs on it as to make sure there was no contention.

I am currently working with VMWare on this also but they have not been helpful so far. I am also working with Dell Pro Support but they are pointing the finger at VMWare. I will bounce your theories off of both of them to see what they think.

We did attempt to disable all power saving features in the BIOS on the E5-2670 but that did not seem to help.

There is also an option in VMWare to assign both number of virtual sockets and number of cores per socket. I played with them both but it did not change the test result. 2 virtual sockets and 1 core per socket OR 1 virtual socket and 2 cores per socket both give me results around 600.

Is there anything about the single threaded test that could explain the difference or is it a pretty good representation of real-world single threaded performance?

[David (PassMark)]

The numbers you posted seem to completely contradict your previous explanation of the problem.

You said, "If I change the VM config to just have 1 vCPU .... double the score on the E5-2670". But the numbers say the opposite.

We have a fairly comprehensive list of results for the single threaded test here,
http://www.cpubenchmark.net/singleThread.html

We are not aware of any issue in the test itself that would cause this & the list seems to be more or less what I would expect in terms of rankings.

The 661 result is bad. This is around the performance of a Pentium 4

[ckbrou]

Sorry about that - that was a typo on my part. Here are the results:

VM w 1CPU X5460: 1240
VM w 2 CPUs X5460: 1208
VM w 1 CPU E5-2670: 1308
VM w 2 CPUs E5-2670: 661

I have been working with VMWare and it is looking like when 2 vCPUs are assigned to the VM the processor does fail to ramp up fully as you suspected. Now we just need to find out why. I will post back here when I find out.

[ckbrou]

So it was related to power saving features...
The problem ended up being with the BIOS settings. I had to clear NVRAM for it to start working right again. Clearing NVRAM and setting the System Profile to “Performance” (which disables C1E and C States) took care of it. Single threaded CPU performance is now consistently about 15 -20% faster than with the older X5460 processor and over twice as fast as the E5-2670 when it had the C1E and C States enabled.

The single threaded result is now 1416.

[David (PassMark)]

Thanks for posting the solution.
Glad you sorted it out (and that our software helped in the process).