Does EVC mode affect CPU performance?

[biokovo]

We have vSphere4 with 3 server: A and B servers with 2 Intel 5440 CU (Penrym), and C server with 2 Intel 7320 CPU (Merom).

On each of them are a few VMs with virtual hardware ver.7.

We create DRS cluster, set EVC mode Intel Xeon Core 2 (server A and B support Intel Xeon 45nm Core 2 EVC mode but server C doesn't) and put server C at cluster.

When try to put server A and B with running VMs there is an error because running VMs(?!)

We shut down VMs on server A and B, put A and B at cluster and power on VMs without any problem.

My questions are:

1. Why I couldn't put server A and B with running VMs in cluster?

2. Is EVC mode affect CPU performance, e.g. is it better to set Intel Xeon 45nm Core 2 than Intel Xeon Core 2?

3. I read some documentation about overriding mask on VMs. Is it a good option if you need that?

Thanks

[AntonVZhbankov]

>1. Why I couldn't put server A and B with running VMs in cluster?

Because running VMs already see CPU as Core 2 45nm. So you have to power these VMs down, move host to cluster and power VMs on.

>2. Is EVC mode affect CPU performance, e.g. is it better to set Intel Xeon 45nm Core 2 than Intel Xeon Core 2?

Actually only if you're using VM for multimedia software like 3d rendering. EVC will disable such features as latest SSE instructions. But in general you won't see any performance difference between standalone host and host in EVC mode.

>3. I read some documentation about overriding mask on VMs. Is it a good option if you need that?

EVC is a way to override CPUID masks automatically, without need to configure long bit masks for each VM manually.

---

MCSA, MCTS, VCP, VMware vExpert '2009

http://blog.vadmin.ru

[Matt_B1]

We have Intel® Xeon® 5600

Series blades that would fit into this EVC baseline, Intel® 32nm Core™ i7

(Westmere). We also have older Intel® Xeon® 7300

Series servers that can only go in this EVC baseline, Intel® Core™2 (Merom). If we were to use the lowest baseline to allow vMotion across all of these hosts, we would lose these features (SSE4.1, SSE4.2, POPCOUNT, AES, PCLMULQDQ and RDTSCP). Does anyone have proof from testing or some evidence that explains our 5600 series blades will not have slightly degraded performance by masking out these features? We are not aware if any VMs in our environment take advantage of the CPU features nor do we know how to check for that.

[jonowebster]

I've been testing this this last week and it very much does make a HUGE diference. We had a number of 'old' Xeon E5420's running with E5520 and also E5620's in the same cluster at the EVC Core 2. Level.

Taking a single node with E5620 CPU's and placing it in a cluster by itself at the core i7 EVC level saw a near two fold performance increase. Windows VM's have to be fully powered down, powered up and then restarted again after they pickup the hardware change (it will automatically prompt for the reboot) , then watch them grow legs and power away, its quite insane to watch.

Happy for someone else to test this too, but simply restarting the VM's twice won't show any difference , you have to power them down and power them up for detect the EVC change.

[Matt_B1]

It is good to know that you were able to actually see a performance increase due to upgrading the EVC level. Were the VMs you tested running on the E5620s before and after the test? I just took a bunch of downtime to enable EVC so I am not excited about requesting more to increase the baseline and have the VMs actually realize the difference.

Has anyone else been able to actually see a performance increase when upgrading the EVC baseline?

[jonowebster]

The VM's were already on the 5620's. I may have been a little over optimistic in the increased performance, I'm not sure if this was an across the board increase but rather when doing CPU performance testing / number crunching. One customer did comment that 'things feel more snappy" ,.

Ultimatly the application will need to be making use of the added CPU features to notice a difference, for us we went from EVC Core 2 to i7 so a bunch of new features were added.

I'd be interested to see what your results are like, maybe just remove one host from your cluster, clreate new cluster with the highest EVC level and add the host to it. then fire up a test VM so you don't affect anyone's uptime.

VM's I've tested on were 2008 R2 std