I'm surprised. I definitely recall having that conversation before about never exceeding the physical limits of the hypervisor. I'm pretty sure we have test cases against that as well.
--Alex > -----Original Message----- > From: Edison Su [mailto:edison...@citrix.com] > Sent: Wednesday, May 23, 2012 1:39 PM > To: cloudstack-dev@incubator.apache.org > Subject: RE: Vmware CPU Cap > > I think it's a general issue for all the hypervisors we supported. > > > -----Original Message----- > > From: Alex Huang [mailto:alex.hu...@citrix.com] > > Sent: Wednesday, May 23, 2012 1:32 PM > > To: cloudstack-dev@incubator.apache.org > > Subject: RE: Vmware CPU Cap > > > > That's because no one wrote a host allocator for VmWare. > > > > The relationship between DeploymentPlanner and HostAllocator and > > StoragePoolAllocator is as follows: > > > > DeploymentPlanner deals with heuristics and affinity rules of placing > > a VM. Once it determines a set of hosts that matches, it then asks > > the HostAllocator to work on the limitations of the type of hypervisor. > > There was never a HostAllocator written for VmWare. It just uses the > > one that was written originally for XenServer. Hence the bug. > > > > There's similar differentiation for DeploymentPlanner and > > StoragePoolAllocator and similar bugs exists, especially if someone > > adds a completely new type of storage pool but decides to just reuse > > the current set of StoragePoolAllocators. > > > > --Alex > > > > > -----Original Message----- > > > From: Edison Su [mailto:edison...@citrix.com] > > > Sent: Wednesday, May 23, 2012 1:24 PM > > > To: cloudstack-dev@incubator.apache.org; 'tam...@veber.co.uk' > > > Subject: RE: Vmware CPU Cap > > > > > > Thanks for your input, I find another bug in cloudstack... > > > Let's dive little bit deeper into how the CPU allocation works. > > > VMware(the same for other hypervisors(Xen/KVM)) uses proportional > > share > > > based scheduling algorithm([1],[2]). It means "If a virtual machine > > has twice > > > as many shares of a resource as another virtual machine, it is > > entitled to > > > consume twice as much of that resource when these two virtual > > machines > > > are competing for resources." > > > There are two questions: > > > 1. Where is the share coming from? In Cloudstack, the share is > > calculated > > > from (CPU MHz * CPU cores). > > > 2. How the share of a VM is mapped to the physical CPU by the > > hypervisor? > > > Of cause, it depends on the hypervisor implementation, but there are > > some > > > general rules that we need to follow, as we are living in the real > > world:) > > > Rule 1: num of cpu core of a VM should be <= num of cpu core of > > the > > > hypervisor host. It doesn't make sense to running a 12 core VM on a > > single > > > core host, even some hypervisors, such as KVM, can do that. > > Hypervisor will > > > be busy at scheduling VCPU(the VCPU context switch is much heavier > > than > > > process context switch), thus bad performance for the VM. > > > Rule 2: The frequency of a VCPU should be <= the frequency of the > > host > > > hypervisor host. Can you run an one core * 12Ghz VM on a 3Ghz * 4 > > core > > > physical CPU? Nope, in an unit time, the max freq of VCPU can get is > > the max > > > freq of physical cpu core, that's the physical LAW. > > > Rule 3: the total share of VMs <= total share of host > > hypervisor(in case of > > > no CPU overcommit). > > > > > > Currently, in cloudstack host allocator, only rule 3 is taken into > > consideration. > > > > > > [1] www.cs.uiuc.edu/class/sp10/cs423/lectures/17-VMM-resource.pdf > > > [2] > > > > http://www.vmware.com/pdf/vsphere4/r40/vsp_40_resource_mgmt.pdf > > > > > > > -----Original Message----- > > > > From: Tamas Monos [mailto:tam...@veber.co.uk] > > > > Sent: Tuesday, May 22, 2012 5:41 PM > > > > To: cloudstack-dev@incubator.apache.org > > > > Subject: RE: Vmware CPU Cap > > > > > > > > Hi, > > > > > > > > I still don't think this is an issue as the CPU Mhz limit and the > > > > number of cores are independent. > > > > CPU manufacturers sell 2,4,6 cores at 3Ghz and not 6,12,18Ghz CPUs. > > > > > > > > So I think it is good how it works but the "number_of_cores*Mhz" > > while > > > > allocating should not multiply so that is the bug :) What vmware > > does > > > > with multiplying the cores with the core speed is bad as I can't > > have > > > > a 1vCPU VM at 12Ghz on a 3Ghz quad core if you know where I'm > > coming > > > > from.... > > > > > > > > Regards > > > > > > > > -----Original Message----- > > > > From: Tamas Monos [mailto:tam...@veber.co.uk] > > > > Sent: 23 May 2012 01:27 > > > > To: cloudstack-dev@incubator.apache.org > > > > Subject: RE: Vmware CPU Cap > > > > > > > > Hi, > > > > > > > > I have confused myself too because if I have a look the database > > and > > > > dump the service_offerings table the limit for me is set to > > > > 2000,3000,4000 and not 1000. > > > > So a Mhz limit set to 4000 and 4 cores will end up as a quad core > > box > > > > at 4Ghz. I remember now I had to set CPU overprovisioning to 4 as > > > > CloudStack took away 4x4Ghz=16Ghz of the available CPU. > > > > So Cloudstack sets the VM CPU Mhz limit what the actual limit is > > set > > > > to in the offering but does not multiply the Mhz limit by the > > number > > > > of cores when setting the limit on vmware. However takes away > > > > "number_of_cores*Mhz limit" from the available CPU capacity when > > > > allocating. > > > > > > > > I'm getting confused myself so I'm not sure if this is bug now or > > not > > > > either :) I'm using 3.0.3 > > > > > > > > Regards > > > > > > > > -----Original Message----- > > > > From: Edison Su [mailto:edison...@citrix.com] > > > > Sent: 23 May 2012 00:32 > > > > To: cloudstack-dev@incubator.apache.org; Tamas Monos > > > > Subject: RE: Vmware CPU Cap > > > > > > > > But in Diego'case, the limit is set as 1000Mhz, while his service > > > > offering is 1000Mhz * 2 cores. > > > > Which version of cloudstack are you using? Maybe it's a regression. > > > > > > > > > -----Original Message----- > > > > > From: Tamas Monos [mailto:tam...@veber.co.uk] > > > > > Sent: Tuesday, May 22, 2012 4:26 PM > > > > > To: cloudstack-dev@incubator.apache.org > > > > > Subject: RE: Vmware CPU Cap > > > > > > > > > > Hi, > > > > > > > > > > I have a service offering with 1000Mhz limit and 4 cpu cores. > > That > > > > > sums up to 4000Mhz. > > > > > Cloudstack sets the limit to 4Ghz on the virtual machine and > > > > > when > > I > > > > > put load on it vmware balances the load between the 4 cores > > allowing > > > > > them to use 1000Mhz each. > > > > > I do not see any bugs here. > > > > > > > > > > Apologies if the "meant per CPU core" was incorrect. What I > > > > > meant > > is > > > > > described above. > > > > > > > > > > Regards > > > > > > > > > > -----Original Message----- > > > > > From: Edison Su [mailto:edison...@citrix.com] > > > > > Sent: 23 May 2012 00:05 > > > > > To: cloudstack-dev@incubator.apache.org > > > > > Subject: RE: Vmware CPU Cap > > > > > > > > > > Nope, from the document, the limit is set on whole vm: > > > > > CPU Limits > > > > > > > > > > When a CPU Limit is set on a virtual machine resource settings, > > the > > > > > virtual machine is deliberately held from being scheduled to a > > PCPU > > > > > when it has used up its allocated CPU resource. This happens > > > > > regardless of the CPU utilization. If the limit is set to > > > > > 500MHz, the virtual machine is descheduled from the PCPU and has > > > > > to wait before > > > > it > > > > > is allowed to be scheduled again. As such, the virtual machine > > might > > > > > experience performance degradation. > > > > > > > > > > Note: For an SMP virtual machine, the sum of all vCPUs cannot > > exceed > > > > > the specified limit. For example, 4 vCPU virtual machine with a > > > > > limit of 1200MHz and equal load among vCPUs would result in a > > > > > max > > of > > > > > 300MHz per vCPU. > > > > > > > > > > > > > > > > > > http://kb.vmware.com/selfservice/documentLinkInt.do?micrositeID=&pop > > > u > > > p > > > > > = > > > > > true&languageId=&externalID=1033115 > > > > > > > > > > > -----Original Message----- > > > > > > From: Tamas Monos [mailto:tam...@veber.co.uk] > > > > > > Sent: Tuesday, May 22, 2012 3:43 PM > > > > > > To: cloudstack-dev@incubator.apache.org > > > > > > Subject: RE: Vmware CPU Cap > > > > > > > > > > > > Hi, > > > > > > > > > > > > No I don't think this is a bug. When you set 1000Mhz as CPU > > > > > > cap > > > > that > > > > > > is meant per core. vmWare will limit each CPU core to 1000Mhz. > > > > > > As you gave 2 CPU cores that is 2000Mhz effective. That is how > > > > > > vmware works. > > > > > > > > > > > > I have setup my offerings all to 1000Mhz as speed and just > > > > > > increasing the number of cores. > > > > > > 1 core ends up being 1x1000Mhz > > > > > > 2 core ends up being 2x1000Mhz=2000Mhz ... > > > > > > ... > > > > > > > > > > > > I'm actually using it in production and works quite well as > > > > > Cloudstack > > > > > > "allocates" 4000Mhz when I'm using 4x1000Mhz cores and vmware > > > > > cleverly > > > > > > balances between the cores as you put load on it and not > > letting > > > > any > > > > > > of the cores above the set limit of 1000Mhz. > > > > > > > > > > > > > > > > > > Regards > > > > > > > > > > > > -----Original Message----- > > > > > > From: Diego Spinola Castro [mailto:spinolacas...@gmail.com] > > > > > > Sent: 22 May 2012 19:34 > > > > > > To: cloudstack-dev@incubator.apache.org > > > > > > Subject: Re: Vmware CPU Cap > > > > > > > > > > > > I forgot the link: > > > > > > http://imageshack.us/photo/my-images/848/cpulimit.png/ > > > > > > > > > > > > 2012/5/22 Diego Spinola Castro <spinolacas...@gmail.com> > > > > > > > > > > > > > I believe that is a bug with cpu cap and vmware. > > > > > > > > > > > > > > To reproduce: > > > > > > > > > > > > > > Create a offering with 2 cores and 1000mhz. > > > > > > > Enable CPU CAP. > > > > > > > > > > > > > > After created instance , cs create a vm with 2 cores and > > 1000mhz > > > > > > > of > > > > > > limit. > > > > > > > > > > > > > > I don't know for sure if is a bug, but vmware gives 1000mhz > > > > shared > > > > > > > with cores. > > > > > > > > > > > > > > > > > > > > > > > > > > > > Diego > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >
