INTEL vs. AMD with ESX 4.1

[Buck1967]

I've been chasing a performance issue the past few days and ran into a weird issue. I'm running a win2008R2 image on a ESX 4.1/HP ProLiant BL465c G7 blade. I could not get anywhere close to the performance that a similar physical got (3xslower). After emptying my bag of tricks to resolve the issue, I decided to try it on a ProLiant BL460c G7 blade. Without any modifications to the image it immediately ran in par with the physical. Has anyone else seen anything similar? I'm new to this particular infrastructure, which has all AMD based ESX hosts. The prior infrastructures I supported were Dell servers running on INTEL. is there something simple that I might be overlooking?  I did verify virtualization support was enabled in the BIOS. Is there anything else that I may need to verify? I can’t believe that AMD procs run that much differently than the AMD.

Buck

[JohnADCO]

I can see no reason that would really be.    We run a datacenter of all Intel and a separate data center with all AMD and we have moved lots of machine back and fourth.

Do all other machines seem to behave on the 65??   Or was this a first sort of test VM?

I'd be hitting HP support up big time.   When I have trouble shot HP blade VM performance issues?  It's almost always ends up being the dang flex network or what ever they call it, configuration and /or driver issues with it.

[admin]

Do you have any vmware.log files from the Intel hosts?

[Buck1967]

I'm not sure why you would need INTEL logs, it is the one process fine,,, I might be missing something here,

Here is a brief summary of the Hardware and Software in the environment along with the testing and noted variables.  The information provided below as I fell strongly this is not specifically an AMD Issue rather perhaps a specific configuration variable we’ve overlooked somewhere.

Problem Statement

Issue appears to be around application writes across the network to NAS Devices, which is roughly 3 times faster on the Intel, I just move the image between the systems for the test.  Note, these systems are isolated within the same blade chassis to alleviate to much of a difference in supporting infrastructure configurations.  4 specific applications have currently been identified to be impacted by this behavior: Relativity, Rhinosoft Serv-U, Creatfile “Test Tool”, LANSpeed 2.0 test.

[admin]

Buck1967 wrote:

I'm not sure why you would need INTEL logs, it is the one process fine,,, I might be missing something here,

I'm hoping your performance issue is related to monitor mode.  Both of the blades you mentioned have CPUs that are capable of hardware-assisted CPU virtualization with hardware-assisted MMU virtualization.  Usually, that's the monitor mode that provides the best performance for a win2008 guest.  Hence, that will be the default monitor mode if you haven't selected a preferred monitor mode.  However, on the Intel processors, there are a variety of reasons that hardware-assisted virtualization might not be available, and we would fall back to binary translation (assuming that this is a 32-bit version of win2008 we're talking about).  The vmware.log file will reveal what monitor mode was used.  This could be a case where binary translation outperforms hardware-assisted virtualization.

Then again, this might not have anything to do with monitor mode at all.

[RParker]

Buck1967 wrote:

I've been chasing a performance issue the past few days and ran into a weird issue. I'm running a win2008R2 image on a ESX 4.1/HP ProLiant BL465c G7 blade. I could not get anywhere close to the performance that a similar physical got (3xslower). After emptying my bag of tricks to resolve the issue, I decided to try it on a ProLiant BL460c G7 blade. Without any modifications to the image it immediately ran in par with the physical. Has anyone else seen anything similar? I'm new to this particular infrastructure, which has all AMD based ESX hosts. The prior infrastructures I supported were Dell servers running on INTEL. is there something simple that I might be overlooking?  I did verify virtualization support was enabled in the BIOS. Is there anything else that I may need to verify? I can’t believe that AMD procs run that much differently than the AMD.

Buck

Well SOMETHING must be different, because AMD cannot hang with Intel.  Google it if you want.. ALL day long, Intel runs faster than AMD in every case.  We have different Intel machines and THEY have slight differences, like IBM 3950 right slightly better than Dell, however they are quad core.. Identical specs.  With the 6 core machines the Dells are dramatically improved.. same Intel, but we have AMD machines for production, running Hyper-V and ESX, and Intel just runs better, this is the first I have heard anyone say AMD is faster.. but something in your environment is different.. not JUST the CPU....

At any rate with Virtualization TECHNICALLY there should be ZERO difference.  It should be completely transparent the the VM's and the virtualization, only thing you MAY see is better integration with VM Ware, but definately ESX on Intel / AMD shouldn't be any difference with respect to how it behaves, performance is a different matter.. Intel is faster, not AMD.

[Buck1967]

Sorry if I did not make myself clear... INTEL is by far much faster. In this case the performance on the AMD ESX host is so much worse, I think there has to be some type of driver issue. I have the ServerEngines NC553i 10G interfaces. I've been going through trying to get all the firmware and drivers updated.  I also came accoss a post to disable TCP offloading. That didn't help either. I think I'm going to open a ticket with HP.

[Buck1967]

I have determined that the HP Power Profile setting was the issue causing our performance issues.  The HP BL465 blades (AMD based) come with a default power profile of "Balanced"where as the HP BL460's default is "Maximum Performance". After changing the power profile setting for the BL465 (AMD based) the performance now appears to be inline the INTEL blade. Just to make myself clear, the INTEL is still faster, but now the AMD is in the range of only 15-20% slower rather then to huge difference I had been seeing. I'm not sure why HP would have different default power settings for these systems. I have not testing fully on a system other than one running ESX 4.1, so maybe Windows and/or Linux don't suffer the same from this setting.

BUCK