The time increased due to the newly added dedicated resources feature. During regular VM deployment, all dedicated resources are put in avoid list so that they are not considered for deployment.
Now the way to compute the list of dedicated resources is not optimal and performance deteriorates in an environment having lot of pods, clusters and hosts as the logic is to query db. for each suc resource.
The fix is to optimize the logic not to loop through all resources but get the list of each resource type in a single query.
Issues:
In Implicit planner resource usage is fixed to "Dedicated". It should be Dedicated/Shared depending upon the Implict Planner strict/preferred modes and hosts availability.
Fixed:
Issue is fixed by determining the resource usage to be "Dedicated/Shared" depending upon the Implicit strict/preferred mode and the hosts availability for the planner.
1. A new implicit planner which extends the functionality provided by FirstFitPlanner.
2. Implicit planner can be used in either strict or preferred mode. In strict mode it tries to deploy a vm of a given account on a host on which vms of the account are already running. If no such host is found it'll search for an empty host to service the request. Otherwise the deploy vm request fails.
3. In preferred mode, if a host which is running vms of the account or an empty host isn't found, the planner then tries to deploy on any other host provided it isn't running implicitly dedicated strict vms of any other account.
4. Updated the createServiceOffering api to configure the details for the planner that the service offering is using.
5. Made db changes to store the service offering details for the planner.
6. Unit tests for testing the implicit planner functionality.
7. Marvin test for validating the functionality.
- Changes merged from planner_reserve branch
- Exposing deploymentplanner as an optional parameter while creating a service offering
- changes to DeploymentPlanningManagerImpl to make sure host reserve-release happens between conflicting planner usages.
Reviewed-by: Sheng Yang
Changes:
- Do not check if allocation_state is 'Enabled' in planner if the caller is Root Admin.
- This should let Root Admin create a VM in a disabled Zone.
Changes:
To migrate systems using 'use.user.concentrated.pod.allocation' as true and 'vm.allocation.algorithm' as true, we need to
add following changes:
- There will be 5 values to 'vm.allocation.algorithm': 'random', 'firstfit', 'userdispersing', 'userconcentratedpod_random', 'userconcentratedpod_firstfit'
- 'userconcentratedpod_random' means we apply user concentration to pods and clusters. To hosts and pools we use random ordering.
- 'userconcentratedpod_firstfit' means we apply user concentration to pods and clusters. To hosts and pools we use firstfit ordering.
Changes:
- Added a two new deployment planners 'UserDispersingPlanner' and 'UserConcentratedPodPlanner' to the DeploymentPlanners
- Planner can be chosen by setting the global config variable 'vm.allocation.algorithm' to either of the following values:
('random', 'firstfit', 'userdispersing', 'userconcentratedpod')
- By default, the value is 'random'. When the value is 'random', FirstFitPlanner is invoked as before that shuffles the resource lists.
- Now Admin can choose whether the deployment heuristic should be applied starting at cluster or pod level. This can be done by using the
global config variable 'apply.allocation.algorithm.to.pods' which is false by default. Thus by default as earlier, planner starts at clusters directly.
'UserConcentratedPodPlanner' changes:
- Earlier to 3.0, FirstFitPlanner used to reorder the clusters in case this heuristic was chosen.
- Now this is done by a separate planner and is applied only when 'vm.allocation.algorithm' is set to this planner
- It reorders the capacity based clusters/pods such that those pods having more number of Running Vms for the given account are tried first.
- Note that this userconcentration is applied only to pods and clusters. Not to hosts or storagepools within a cluster.
'UserDispersingPlanner' changes:
- 'UserDispersingPlanner' reorders the capacity ordered pods and clusters based on number of 'Running' VMs for the given account in ascending order. Aim is to choose thodes pods/clusters first which have less number of Running VMs for the given account
- Admin can provide weights to capacity and user dispersion so that both parameters get considered in reordering the pods/clusters. This can be done by setting
the global config parameter 'vm.user.dispersion.weight'. Default value is 1. Thus if this planner is chosen, by default, ordering will be done only by number of Running Vms, unless the weight is changed.
- HostAlllocators and StoragePoolAllocators also reorder the hosts and pools by ascending order of number of Running VMS/ Ready Volumes respectively for the given account. Thus try to choose that host or pool within a cluster with less number of VMs for the account.
- Added a new flag 'allocation_state' to zone,pod,cluster and host
- The possible values for this flag are 'Enabled' or 'Disabled'
- When a new zone,pod,cluster or host is added, allocation_state is 'Disabled' by default.
- For existing zone,pod,cluster or host, the state is 'Enabled'.
- All Add/Update/List commands for each of zone,pod,cluster or host can now take a new parameter 'allocationstate'
- If 'allocation_state' is 'Disabled', Allocators skip that zone or pod or cluster or pod.
- For a root admin, ListZones lists all zones including the 'Disabled' zones. But for any other user, the 'Disabled' zones are not included in the response.
- For any usecase that creates/deploys/adds/registers a resource and takes in zone as parameter, now we check if the Zone is 'Disabled'. If yes then the operation cannot be performed by a user other than root-admin. Add volume, snapshot, templates are examples of this usecase.
- To enable the root admin to test a particular pod/cluster/host, deployVM command takes in 'host_id' parameter that can be passed in only by root admin.
If this parameter is passed in by the admin, allocators do not search for hosts and use that host only. StoragePools are searched in the cluster of that host.
If VM cannot be deployed to that host, allocators and deployVM fails without retrying
Bug 7723 - merge or re-write host tagging into master / 2.2
Bug 7627 - Need more logging for Allocators
Bug 8317 - Add better resource allocation failure messages
Changes for Deployment Planner to use host and storagePool allocators to find deployment destination.
Also has the changes for host tag feature.
Improved the logging for allocators.
Nitin Mehta
Koushik Das
Rajesh Battala
Devdeep Singh
Prachi Damle
Rohit Yadav
David Nalley
frank
Alena Prokharchyk
alena
Sheng Yang
Alex Huang
Frank
edison
root
Manuel Amador (Rudd-O)