Friday, May 8, 2020

vSphere Distributed Resource Scheduling - DRS

 DRS is a core technology for resource management in a vSphere cluster. It has been around since ESX3 and it's a battle proven feature without vSphere clusters would not look the same. But what it actually does?
At a high level, it enables to use the resources of ESXi hosts in a cluster as an aggregated pool of resources. Drilling a bit into what it does we'll see that:

  • it provides virtual machine admission control - are there enough resources in the cluster to power on a VM
  • it provides initial placement of a VM - what is the most appropriate host to power on the VM
  • it is responsible for resource pools - quantifiable and aggregated resources to be consumed by a VM or group of VMs
  • it is responsible for resource allocation to VMs or resource pools using shares, reservation and limits
  • it balances the load in the cluster 

vSphere 7 comes with an important change in the logic DRS uses. Until vSphere 7, DRS would try to balance the load looking at the cluster. If a host was overloaded at some point in time, it would try to balance it by migrating VMs to less utilized hosts. Checking cycle was 5 minutes. Starting with vSphere 7 the focus has shifted to VM. DRS calculates a per VM score called virtual machine happiness. Looking at the VM and running every minute, provides a better way of load balancing and ensuring placement of VMs.

Let's look at some of the features in DRS as they appear in the UI. As you can see above at the cluster level you can see the score of the cluster (an average of the scores of each VM) as well the score buckets for VMs. All my VMs are happy in the 80-100%, meaning they have all the resources they require. Going to VM view, we'll see the individual VM scores as well as some of the monitored metrics such as CPU % ready, swapped or ballooned memory:

DRS is enabled at cluster level. Once enabled, four tabs get activated.

Automation tab
The first choice is how much freedom you give to DRS: Automation Level.

There are 3 levels you can choose from

  • manual - generates recommendations for initial placement and migrations. But you have to actually apply the recommendations. Hence it is manual intervention every time.Very good when you need to do some troubleshooting. 
  • partially automated - initial placement of VMs is done by DRS, but migrations are kept at recommendation level. 
  • fully automated - DRS will take care of both initial placement and migrations
Once you have decided which automation level to use, you will choose the threshold for which migrations should be made. The slider is scaled from conservative to aggressive. DRS looks at an imbalance in the cluster the five levels on the slider determine how big that imbalance can be. A conservative setting will not generate migration recommendation for load balancing. An aggressive setting will calculate a very small imbalance threshold.  This translates from almost no migrations (except for specific cases like putting a host into maintenance mode) to a lot of migrations. 

Predictive DRS has been introduced with vSphere 6.5 and it utilizes metrics from vRealize Operations Manager to balance predicted cluster load and workload spikes. 

Virtual Machine Automation enables VM level override of DRS and HA settings. When enabled, you can specify at Cluster - Configure - VM overrides the VMs for which you would change the default settings like having them excluded from migration recommendations:

Additional Options tab

VM Distribution instructs DRS to try and evenly distribute the VMs on hosts. It is a soft limit that will not be enforced over migration recommendations. 

CPU Over-Commitment enforces the defined ratio of vCPU/core. When enabled, DRS will not allow to power on VMs if the ratio is overpasses. This enables to keep some clusters in the realm of performance. The max value is 32, this being the maximum vCPU/core for vSphere 7. 

Scalable shares is a new feature introduced in vSphere 7. You can find very good articles here and here. In a nutshell, scalable shares takes care that the shares allocated to a VM are actually taking into consideration the share priority (high, normal, low) and avoids situations where VMs in resource pools with lower priority can get more resources than VMs in resource pools with higher priority. This situation is called resource pool priority-pie paradox.

Power Management tab

When activated, Distributed Power Management (DPM) looks at the cluster utilization and consolidates VMs on hosts based in order to power off hosts and save energy. For more details, you may look at this article 

Advanced options tab

The tab displays advanced options that have been set for DRS through the UI or manually.

This has been a small introduction to DRS as it looks now in vSphere 7. There are a lot of features and details that have been barely touched or not touched at all. For a deep dive, I recommend the famous Clustering Deep Dive book although I am waiting for an updated version.

No comments: