Anand_Arvind
Contributor
Contributor

improving performance of a single VM past a PassMark CPU Mark of 1000

Our setup includes a Dell M600 blade with Intel E5410 chipset based dual proc quad core processors with 32 GB RAM running VM ESX 3.0.1. This works very well when we run 16-20 VM's on this machine. The VM's are based on Windows 2003 Server 64 bit versions. A challenge we have is we are trying to provide some fast VM's and are using PassMark (www.passmark.com) to quantify CPU performance of the VM's. We are currently not able to push this past a CPU mark of around 1000 while the base machine is rated at close to 70008000 (and we have measured this on another machine). This is with a 4 core VM which is the max I see supported in ESX 3.0.1. Any pointers to help to improve performance on VM's would be appreciated as I need to figure out how to do this with VM's as high end desktops (2 proc-2core) are providing ratings of 20003000 which we need to be able to match with VM's.

I just joined this community so if there are already pointers to answers do send them my way.

0 Kudos
9 Replies
MetronAthenePhi
Contributor
Contributor

Hi Anand.

The 1st thing I would say is try reducing the systems back down to 2 virtual CPUs (or even better 1 if it gives better results). The more Virtual CPUs in a VM can cause contention on the system as no processing can be done for the VM without 4 available processing cores on the hardware, even if only one is actually required by the OS in the VM. As they are coming from 2 proc desktops I presume you dont NEED to have 4 concurrent CPUs in the VM.

If there is contention it should show up as CPU Ready time in the performance statistics you can view in Virtual Center.

Hope that Helps

Phil

0 Kudos
drummonds
Hot Shot
Hot Shot

Hi, Anand,

PassMark is one of our tried-and-true benchmarks for measuring CPU performance in-house here at VMware. We feel confident that its CPU and memory (but not video and storage!) results are good. In fact, check out figure four from the Performance Comparsion of Hypervisors paper and you can see the results that you should expect from the CPU tests in PassMark.

If my memory serves me correctly, PassMark makes use of extra cores. That is, all other things being equal, more cores equals a higher score. So, probably the difference you're seeing in performance is due to the presence of eight cores in your native runs and fewer (1? 2? 4?) in your VM. If you want to get an accurate native-to-virtual comparison, try throttling the core count through boot.ini to measure native performance. Then you can see what the impact of virtualization is.

But, if you're looking to get the highest possible numberone comparable to the eight-core run you're comparing againstyou're going to have to wait for the next version of ESX. ESX Server 3.5 only supports 4-way VMs.

Scott

More information on my blog and on Twitter: http://vpivot.com http://twitter.com/drummonds
0 Kudos
Anand_Arvind
Contributor
Contributor

Thank you for the pointer to the PDF, I did read this earlier in the day before posting to this forum. The paper dealt with relative numbers and did not give me a feel for the absolute numbers that could be achieved.

I am right now not trying to compare the VM performance with the native machine performance to see how fast VM works, but just want to figure out how to get a fast VM - measured by a high PassMark CPU mark. My logic is that if natively the machine with 8 cores (2x4) can support a rating of close to 7000 ~ 8000 with this test, then a VM on this with 4 cores can I get to a rating of around 2000 ~ 3000? This is lower than something like 3500 accounting for the tax of being in a VM and if I get close to this I am going to be in good shape. The need arose because my performance team has machines in this range which work well for them and to get them to move to VM's I need my VM's performing at this level when necessary (these are usually 2 cpu x 2 core machines). We already run a setup with 6 of these servers in a cluster supporting around 100 VM's (use SAN for storage). I have 10+ more blades like this which run without VM's on them and if VM works want to move more towards VM but need to be able to provide fast VM's to my team when necessary.

I will ask my team to run a matrix of tests tomorrow to see what I get by tweaking parameters and see if we make any headway on this. If there are any other pointers that exist that we can tweak do let me know.

Anand

0 Kudos
drummonds
Hot Shot
Hot Shot

I'm not sure if the PassMark scoring system produces linear scaling like that. I suppose that we could test if your theory is true by running 4-way native and 8-way native to see if the change is 2X. I can promise you that the last desktop benchmark I developed (SYSmark) did not have linear scaling as cores were increased. But I agree that superlinear scaling (such that produces 2000 on a 4-way and 8000 on an 8-way) is unlikely, too. But, I simply cannot say.

I can't comment on what numbers your system should expect. We don't run PassMark on that many different systems in here. I can say that when doing a legitimate native/virtual comparison, PassMark's CPU tests should report the VM running at about 90% of native. That being the case, I would expect your system to score 11% higher when you remove VI3 from the equation.

Scott

More information on my blog and on Twitter: http://vpivot.com http://twitter.com/drummonds
0 Kudos
Anand_Arvind
Contributor
Contributor

Thanks for the inputs so far. We did run some tests and I stand corrected on my initial post as we were able to get to some pretty nice numbers with the 4 core machines. Earlier it looks like we had some issues in running the tests. For everyone's benefit here are the results we got - which scaled linearly with cores till 4 the limit with ESX. Thanks for the answers and support so far. Now we will be looking to test the web and database tests with our application and see how this compares with physical machines, and any limits in these. All tests run on a Dell M600 blade with E5410 chip which has 8 cores running at 2.33 GHz.

S. No

RAM

CPU Mark (x32)

CPU Mark (x64)

CPU - Resource Allocation

4 core

2 core

1 Core

4 core

2 core

1 Core

Share

Reservation

Limit

1

4 GB

2947.2

1530.4

803

3679.6

2027.6

1018.9

Normal

No

n x 2.33

2

4 GB

3000.1

1545.2

806

3653.3

2052.4

1014.6

High

No

n x 2.33

3

4 GB

2966.1

1588.7

796.4

3666.8

2046.4

1012.6

low

No

n x 2.33

4

4 GB

2973

1550.5

794.9

3770.6

2023.3

1005.1

Normal

No

Unlimited

5

4 GB

3087.1

1535.4

804.6

3630.9

2036

1033.1

High

No

Unlimited

6

4 GB

3013.5

1548.1

792.3

3505.1

2080.5

1016.4

low

No

Unlimited

0 Kudos
Anand_Arvind
Contributor
Contributor

The table column headers are a bit mixed but but read S.No,RAM, Win2k3 32 bit CPU Mark for 4 core, 2 core, 1 core, Win2k3 64 bit CPU Mark for 4 core, 2 core 1 core, CPU resource allocation Share, Reservation, Limit. The cut/paste from Outlook of an Excel table did not work perfectly.

0 Kudos
meistermn
Expert
Expert

I did it with open office: Is it right now?

S. No

RAM

CPU Mark (x32)



CPU Mark (x64)



CPU - Resource Allocation





4 core

2 core

1 Core

4 core

2 core

1 Core

Share

Reservation

Limit

1

4 GB

2947.2

1530.4

803

3679.6

2027.6

1018.9

Normal

No

n x 2.33

2

4 GB

3000.1

1545.2

806

3653.3

2052.4

1014.6

High

No

n x 2.33

3

4 GB

2966.1

1588.7

796.4

3666.8

2046.4

1012.6

low

No

n x 2.33

4

4 GB

2973

1550.5

794.9

3770.6

2023.3

1005.1

Normal

No

Unlimited

5

4 GB

3087.1

1535.4

804.6

3630.9

2036

1033.1

High

No

Unlimited

6

4 GB

3013.5

1548.1

792.3

3505.1

2080.5

1016.4

low

No

Unlimited

0 Kudos
Anand_Arvind
Contributor
Contributor

Yes and thank you.

You will see that for the 64 bit OS there is a 20% gain in PassMark ratings. We are going to be running some DB/Web tests with our product to really see how this compares to ensure things work well and we address concerns that VM's are slow as the numbers show the VM's can be fast. Since we are connected to a SAN even disk access should be faster than regular disks on desktops. Once I get some data I will post it later though since the tests are using our own product they may not be as good a comparison as PassMark, but they will be real world sustained tests.

0 Kudos
epping
Expert
Expert

Very interesting post

0 Kudos