Relative Weight and vCPU allocation query

Question

Hi all
I've done a lot of reading around this subject mainly starting here:
http://social.technet.microsoft.com/wiki/contents/articles/1234.hyper-v-concepts-vcpu-virtual-processor.aspx
So I know the technical descriptions and some of the background but there is one key area on which I am seeking some clarification if possible.
Most of my Hyper-V experience is with 2008 and 2008R2 so that's what I will be asking about (but it would be useful to know if 2012 changes anything).
I understand the concept of virtual CPUs (vCPU) as given to a VM via the VM settings dialog box. I understand that giving a VM 1 to 4 vCPUs does not give the VM any sort of 1:1 or exclusive right to any of the parent's CPU cores. My understanding is that
(basically) the parent server's cores are formed into a 'pool' and each VM is given a 'time share' of this pool of CPU. I know it is more complex than that due to the way threads may not easily be moved from core to core but let's ignore that!
I understand the 'relative weight' option is used to distribute the 'grunt' from the pool of parent CPU to the VMs such that each VM gets a proportion of the parents CPU-pool as per the ratio of the relative weight scores.
So as an arbitrary, simplified example let's imagine the parent has 2 physical quad core CPUs running at 2.0Ghz. So that's 8 cores running at 2Ghz. So let's imagine the CPU-pool 'grunt' is 16Ghz. (Grunt could be also called 'power', 'horsepower' etc.) I
know it doesn't work out quite like this but this is simplified.
So we add some VMs.
Let's add 4 VMs all with 1 vCPU and all equal 100 relative weight (RW). For the sake of argument all the VMs are 'busy' so that they always need CPU. Let's say each of them shows fairly constant 80% CPU activity in Task Manager within the VM. None of the
VMs are Terminal Servers. Let's also ignore the Hyper-V parent partition's draw on CPU as negligible.
So in this case, if there is contention for the parent CPU, then each of the 4 VMs would be allocated an equal time share of the parent's pool of CPU. So it might be said that on average each VM runs as if it had 4Ghz of 'grunt'. Inside each VM the applications
would see a single vCPU running at 4Ghz. (Bear with me on Ghz, I could have called Ghz 'gruntz' or indeed anything, but I think it helps to talk in terms of Ghz as this is familiar).
Is this correct? Assuming it is then ...
So now I need more 'grunt' to be available to the first VM. So some people may consider two options. Increase the number of vCPU assigned to the VM or increase the Relative Weight. This is where I am confused.
So my question is if the RW is kept constant does increasing the number of vCPU make any more grunt available to a VM or is the grunt simply divided up. For example:
I change VM1 so have 4 vCPU. My understanding is that because RW has not changed then the applications running on VM1 see things change from 1 vCPU running at 4Ghz to 4 vCPUs running at 1Ghz. I surmise this because the RW is what dictates the amount of time
a VM gets on the parent pool of CPU so the VM's allocation (of 4Ghz) must be divided by the number of vCPU.
However I am continually faced with clients asking to allocate more vCPU because they feel this improves the performance of the applications running on the VM, hence I wanted some clarification on whether my understanding is correct or not.
I know that Brian Ehlert, and others, have often stated that it is best to try and give VMs as few vCPUs as possible so that the VM's OS doesn't need to spend time shuffling threads. If this is the case then what is the background or reason for increasing
the allocation of vCPU to a VM? Why add more? When does it have benefit? Is it also required to increase the RW at the same time as adding the vCPU?
It would be great if there was an in-depth article from Microsoft explaining how this all works, in the meantime I'm hoping for some enlightenment from this forum.
Thanks in advance.
Ian

Answer 1

Thanks for your input Ben, that is a very interesting document indeed.
The section that is most pertinent is this one:
<<18.2.3       CPU
Weight
The CPU weight is a relative weight assigned to each of the virtual processors of the partition. Unless otherwise constrained by reserves and caps, the scheduler will attempt to weight the run time of the virtual processors scheduled on
a given logical processor according to their relative weights. Let’s consider the case where three partitions, each with one virtual processor are being scheduled on a single logical processor, their weights are 100, 200 and 700, no reserves or caps are in
effect, and all three of the virtual processors have work to perform (that is, they are not idle). In this case, the fraction of physical CPU capacity provided to the three virtual processors would be approximately
10%, 20%, and 70%.
The CPU weight value is expressed as a decimal value from 1 to 10,000 where 100 (the geometric mean) is the typical value.
By default, a partition’s weight is set to 100.>>
After having read this (and your statement about scheduling processors not VMs) I have had a 'lightbulb moment' where I first understand that the various 'switches' such as Relative Weight are per vCPU and not per VM.
I went back to the online help from Hyper-V and it points to this document:
Performance Tuning Guidelines for Windows Server 2008
http://msdn.microsoft.com/en-us/windows/hardware/gg463394.aspx
In this document there are some useful statements that I was not aware of i.e.
<<Hyper-V in Windows Server 2008 supports up to 16 logical processors and can use all logical processors if the number of active virtual processors matches that of logical processors. This can reduce
the rate of context switching between virtual processors and can yield better performance overall.">>
This is not something that I was aware of. I thought that if you had fewer vCPU than logical CPUs on the host then the vCPU would be shared around the logical CPUs in order to fully utilise the parent CPUs. After reading this above it is obvious that this
is not the case.
And this statement:
<<Virtual Processors
Hyper‑V in Windows Server 2008 supports a maximum of four virtual processors per VM. VMs that have loads that are not CPU intensive should be configured to use one virtual processor. This is because of the additional overhead that is associated
with multiple virtual processors, such as additional synchronization costs in the guest operating system. More CPU-intensive loads should be placed in 2P or 4P VMs if the VM requires more than one CPU of processing under peak load.>>
These two statements lead me to the feeling that it is better to add more vCPU to VMs than fewer if you want to 'make the most of' the CPU you have available on the parent. So going back to my original example of a parent with two quad core CPU (no hyper-threading)
we have 8 logical CPUs available. If we have four VMs that running approximately equal, and busy workloads, we should set each VM to at least 2 vCPU. This would ensure all 8 logical CPUs were used.
First question: What would the potential downside be to setting all four VMs with 4 vCPU which would overcommit the vCPU to logical CPU by a factor of 2. Would this cause scheduling overhead and therefore wasteful of CPU?
Second Question: If I wanted to give a greater share of parent CPU to one of the VMs how would it be best done? Adding more vCPUs to the VM or changing the Relative Weight? Or both? Assuming I was using a Hyper-V 2012 parent that allows more than 4 vCPU
per VM which would be the best way of boosting the performance of one of the VMs:
a) VM1 gets 8 vCPU and each other VM (VM2, VM3 & VM4) each gets 4 vCPU
or
b) VM1 gets 200 Relative Weight and 4 vCPU and each other VM (VM2, VM3 & VM4) each gets 100 Relative Weight and 4 vCPU
Many, many thanks for helping me here.
Ian