Viewaskew
Enthusiast
Enthusiast

EVC - Westmere Vs Sandybridge

Jump to solution

Does anyone know if there is any case to envoke using EVC when combining a cluster of Sandybridge (new HP BLC460 G8s) with the slightly older G7s that use the Westmere set? I

’d like not to have to use EVC and loose some of the features on the new chipsets but that said, Im not totally aware if there is a incompatibility and we use EVC, how much this affects performance.


Thanks

0 Kudos
1 Solution

Accepted Solutions
MKguy
Virtuoso
Virtuoso

You will need to use EVC for migrating between Westmere and Sandy Bridge CPUs. See http://kb.vmware.com/kb/1003212 for a list of changes:

Intel® "Sandy Bridge" Generation

Applies baseline feature set of Intel®  "Sandy Bridge" Generation processors to all hosts in the cluster.  Compared to the Intel® "Westmere" Generation mode, this EVC mode exposes  additional CPU features including AVX and XSAVE.

Note: Intel® "Sandy Bridge" processors that do not support AESNI and PCLMULQDQ cannot be admitted  to EVC modes higher than the Intel® "Nehalem" Generation mode.

So you won't be able to take advantage of AVX and XSAVE instruction sets.

It all comes down to a tradeoff you need to consider: If you want the flexibility to mix clusters/migrate VMs between hosts of different CPU generations, activate EVC using the lowest common denominator, which is Westmere in this case.

The question you're now asking is how refraining from certain new instructions translates to real-world performance loss and this question is impossible to answer. There are some quite specific things like AES-NI (already included in Westmere) that might earn you significant gains in some synthetic benchmarks. But unless you can pinpoint and benchmark if your applications actually make use of and benefit from them, you might as well roll a dice to get your answer.

You can at least be sure it shouldn't be any worse than your current Westmeres, assuming similar #core, clockspeed and caches. Taking internal, transparent optimizations in the CPU architechture into account, it should even be faster out of the box.

So much for today's software behaving like todays CPUs are still "basically overclocked 386's at their core".

I think I'd sacrafice some obscure new CPU instruction over the gained flexibility any day. Unless I would know to really benefit from the new instructions or there is no need for this much flexibility. In my opinion, the last really significant change on the Intel side that warranted an exceptionally large sacrafice of flexibility, was the introduction of the Intel Nehalem architecture. This was over 3 years ago.

-- http://alpacapowered.wordpress.com

View solution in original post

0 Kudos
3 Replies
a_p_
Leadership
Leadership

If you want to be able to live migrate VM's you need to actually make the processor options equal, i.e. use EVC. From what I read so far there's no noticeable performance issue using EVC as the additional processor features may only make a difference in very rare cases anyway.

André

MKguy
Virtuoso
Virtuoso

You will need to use EVC for migrating between Westmere and Sandy Bridge CPUs. See http://kb.vmware.com/kb/1003212 for a list of changes:

Intel® "Sandy Bridge" Generation

Applies baseline feature set of Intel®  "Sandy Bridge" Generation processors to all hosts in the cluster.  Compared to the Intel® "Westmere" Generation mode, this EVC mode exposes  additional CPU features including AVX and XSAVE.

Note: Intel® "Sandy Bridge" processors that do not support AESNI and PCLMULQDQ cannot be admitted  to EVC modes higher than the Intel® "Nehalem" Generation mode.

So you won't be able to take advantage of AVX and XSAVE instruction sets.

It all comes down to a tradeoff you need to consider: If you want the flexibility to mix clusters/migrate VMs between hosts of different CPU generations, activate EVC using the lowest common denominator, which is Westmere in this case.

The question you're now asking is how refraining from certain new instructions translates to real-world performance loss and this question is impossible to answer. There are some quite specific things like AES-NI (already included in Westmere) that might earn you significant gains in some synthetic benchmarks. But unless you can pinpoint and benchmark if your applications actually make use of and benefit from them, you might as well roll a dice to get your answer.

You can at least be sure it shouldn't be any worse than your current Westmeres, assuming similar #core, clockspeed and caches. Taking internal, transparent optimizations in the CPU architechture into account, it should even be faster out of the box.

So much for today's software behaving like todays CPUs are still "basically overclocked 386's at their core".

I think I'd sacrafice some obscure new CPU instruction over the gained flexibility any day. Unless I would know to really benefit from the new instructions or there is no need for this much flexibility. In my opinion, the last really significant change on the Intel side that warranted an exceptionally large sacrafice of flexibility, was the introduction of the Intel Nehalem architecture. This was over 3 years ago.

-- http://alpacapowered.wordpress.com

View solution in original post

0 Kudos
MKguy
Virtuoso
Virtuoso

Btw. among the new vSphere 5.1 whitepapers just released, I noticed this document which compares the impact of running a recent CPU under various older EVC baselines: http://www.vmware.com/resources/techresources/10303 

As you can see, except in the AES encryption case, there is generally very little impact on real application performance, even with rather old, pre-Nehalem baselines. At least the one tested there, so your mileage may vary.

-- http://alpacapowered.wordpress.com
0 Kudos