Page History

Introduction

...

...

This document describes the Cpu/Ram overcommit feature.

...

This feature implements the ram overcommit and allows the ram and cpu overcommit ratios to be specified on a per cluster basis.

Use case

...

...

change the vm density on all the hosts in a given cluster. This can be done by specifying the cpu and ram overcommit ratios.

...

- when combined with dedicated resources, it gets better - with dedicated resources, we may have the capability to tell account A will use cluster X. If this account is paying for "gold" quality of service, perhaps, those clusters would have a ratio of 1. If they are paying for "bronze" QoS, their cluster ratio could be 2. 

Design description

...

...

Admin can give the cpu and ram overcommit ratios at the time of creating a cluster or update the values after creating.

Cloudstack will deploy the vms based the overcommit ratios. If the overcommit ratio of a particular cluster is updated, only the vms deployed hereafter will be deployed based on the updated overcommit ratios, this is ensured by storing overcommit ratio with which the vm got deployed stored in user_vm_details. The overcommit ratios for cluster will be stored in the cluster details table and will be inherited from global setting at the time of creation. Also whenever we add a host we will check of the host has the capabilities to perform the cpu and ram overcommiting. These capabilities will be stored in the db.

Supported Hypervisors

...

...

XenServer
KVM
VMware

Capacity calculations on MS

...

...

Capacity calculation model will be changed to align with the hypervisors calculation. When a vm is deployed with "x" overprovisioing factor we want to guarantee (service offering of vm / x ) during its lifecycle even though the over provisioning changes.

When the cluster overprovisioing factor = x and vms are deployed then

• Total Capacity = (actualHardwareCapacity * x)
• Used Capacity = sum (service offering of each running vm) + sum (service offering of each stopped vm in the skipped.counting.hours)  

...

• Total Capacity = (actualHardwareCapacity * y)unmigrated-wiki-markup
• Used Capacity = \ [sum (service offering of each running vm deployed when factor was x) + sum (service offering of each stopped vm deployed when factor was x in the was x in the skipped.counting.hours)\] * y/x \+ sum + sum (service offering of each running vm deployed when factor was y y ) + sum (service offering of each stopped vm deployed when factor was y in the was y in the skipped.counting.hours)

Ideally you shouldn't change the over-provisioning factor in a cluster with vms running. This is because the existing some of the vms got deployed with the previous factor x. 
Lets say you still want to change the factor. On changing it, both used and total capacity are multiplied by this factor to keep a track of available capacity.

Let's understand the capacity calculation below through an example :-

Cluster – c, 
cpu over provisioning = 1, 
Total cpu = 2GHZ

when we deploy 2VMs of 512Mhz service offering each then 
totalCapacity = 2GHz 
AvailableCapacity = 1GHz
UsedCapacity = 1GHZ

Now change the cpu over provisioning ratio of cluster c to 2
totalCapacity = 4GHz 
AvailableCapacity = 2GHz
UsedCapacity = 2GHZ

Notice the difference in multiplication here. Both used and total capacity are multiplied by this factor. Used Capacity in the new model after changing the factor = (service offering of vm / overcommit it got deployed with) * new overcommit => (1GHZ.5 GHZ/1)*2 + (.5 GHZ/1)*2 => 2GHz
The reason is want to guarantee minimum cpu in (service offering of vm / overcommit it got deployed with) in case of contention. So when a vm is deployed with "x" overprovisioing factor we want to gurantee (service offering of vm / x ) during its lifecycle even though the overprovisioning of cluster is changed.
So the reason we So these vms will get .5Ghz each during contention and therefore available is still 1 Ghz during contention.
The reason to scale the used cpu is to keep track of the "actual" amount of cpu left on the hostfor further vm allocation. Keep the focus on available capacity.

Now if we launch 2 VMs with 1Ghz cpu service offering
totalCapacity = 4GHz 
AvailableCapacity = 0GHz
UsedCapacity = 4GHZ 
Calculation for used capacity for 4vms ((service offering of vm / overcommit it got deployed with) * new overcommit) = 
(512Mhz/1)*2 + (512Mhz/1)*2 + (1Ghz/2)*2 + (1Ghz/2)*2 = 4Ghz

In case of contention first 2 vms (512Mhz service offering) get 512Mhz/1 => .5Ghz each and the next 2 vms (1 Ghz service offering and 2 overprovisioning) also get  (1Ghz/2) = .5Ghz each. So adding up means 2Ghz which is the actual capacity of the host and so there is no more capacity left to accomodate more vms.

now suppose we change the over provisioning to 3 
totalCapacity = 6 GHz 
AvailableCapacity = 0 GHz
UsedCapacity = 6 GHZ
Calculation for used capacity for 4vms ((service offering of vm / overcommit it got deployed with) * new overcommit) = 
(512Mhz/1)*3 +(512Mhz/1)*3 +(1Ghz/2)*3 + (1Ghz/2)*3 = 6Ghz

Now this is assuming, you haven't stopped and started the vms all this while. Say now you stop and start 1 VM = with 512Mhz and another VM = with 1Ghz. The over-provisioning factor ratio changes for these vms to 3 each. Note the denominator in the calculation calculation. 
totalCapacity = 6 GHz 
AvailableCapacity = 1.5 GHz
UsedCapacity = 4.5 GHZ
Calculation for used capacity for 4vms ((service offering of vm / overcommit it got deployed with) * new overcommit) = 
(512Mhz/3)*3 +(512Mhz/1)*3 +(1Ghz/3)*3 + (1Ghz/2)*3 = 4.5 Ghz

All this is done to track the available capacity for further vm allocation. If you track the "actual" capacity left on host = .5Ghz (out of 2Ghz). So now you can still create a vm with 1.5 GHz and cluster over-provisioning = 3 and hypervisor will guarantee 1.5/3 = .5 Ghz during contention.

The upside of new model is we are guaranteeing QOS as (service offering of vm / x ) during its lifecycle vs the old model

Hypervisor Calculation

...

Xenserver 

Deploy vm with service offering ‘s’ and memory overcommit factor ‘f’ and overcommit factor ‘c’   --

• Static Min memory = Dynamic Min memory = service offering / factor ==  (s/f)
• Static Max memory= Dynamic Max memory =  service offering == s
• Each vm ensured Min memory during contention. •
• No overprovisioning means min=max 
• Min memory = Max memory= service offering == s
• CPU weight assigned to vm = (service offering / factor) / host hardware speed == (s/c) / actual_host_speed
• CPU Cap = (serviceOfferingSpeed * serviceOfferingCpus) / actual_host_speed 

Vmware

If vmware.mem.reserve = true 

...

Same model is followed for cpu.

KVM

TBD

DB changes

...

...

• Whenever creating a cluster, MS We will add the cpu and ram overcommit ratios in the cluster_details table. They will be inherited from global settings for cpu over provisioning and for Vmware from memory over provisioning global setting. For othe HVs memory over provisioning will be set as 1.memory over provisioning. 
• In case its an upgrade, existing global values will be carried over to the cluster_details. Do note that memory overprov existed only for vmware and would be carried over during upgrade only for vmware.
• vm_details will be populated with these factors present at the cluster level when the vm is deployed. 
• In case the cluster factor changes, vms factor in vm_details wont change until you stop/start the vm.
• For upgrade vm_details will be populated with the cpu over provisioning and memory over provisioning factors(memory only for vmware) from global setting. For other HVs memory over provisioning will be set as 1.

Caveats

...

What should the behavior be if admin changes the overcommit factor for a cluster that conflicts with the current situation. For example,lets assume Cluster X has an over commit factor of 1.5x for memory and the admin wants to change this to 1x - i.e no overcommit (or changes from 2x to 1.5x) - however, based on the "older" factor, CS might already have assigned more VMs - when the admin reduces the overcommit

...

value  

1. if there is no conflict, there is no issue

...

Note :- The overcommit ratios are dynamically plugged into the capacity calculations. All the capacity calculations is done based on the overcommitted value of capacities. So if the overcommit ratios is decreased the used capacity may go beyond 100%. 
Example:
Overcommit =2 
capacity = 2GB
capacity after overcommit = 4GB.
Now if we deploy 3 VM of 1 GB each 
used =3GB
free = 1GB
used % = 3/4 *100 = 75%
if the overcommit ratio is decreased to 1
used = 3GB
free = -1GB
used % = 3/2 *100 =150% (will generate alerts based on this.)

Alert Generation

...

...

All the alerts are generated based on the global cpu/memory threshold values.

HV prerequisites to use cpu and ram overcommit

...

...

The feature is dependent on the OS type ,Hypervisor capabilities, and some scripts.

...

kvm dose not support automatic adjustment of the guest OS memory dynamically

Note -

• Almost all the hosts have the capability to overcommit, and it is up to the admin to make sure of it. Even if the host is not configured properly, cloudstack will try to set the parameters assuming it has capability.
• As of now cloudstack dose not check for any perquisites. It is the admin's responsibility to provision accordingly.

...

Future Tasks

...

• Keep these factors in the service offering.
• Trigger the capacity recalculate when the overcommit is changed. No need to wait until capacity checker runs.
• Create action event on overcommit change.
• Create action event when capacity recalculate is complete.