As per usual, VMware comes out with some really neat upgrades and new iterations on its previously existing technology. One piece of which caught my eye was the well known distributed resource scheduler, also known as DRS, in this case DRS (1.0).
At this most recent VMworld, the future of DRS, DRS 2.0, was announced. It has been noted that VMware on AWS is already running it, and this new DRS 2.0 may eventually/should eventually find its way to other vSphere distributions. In line with many of the changes we have all seen coming and being introduced by VMware, such as Project Pacific, DRS 2.0 is a enablement that keeps in pace with the needs of modern applications. In a sense, this is the type of innovation consumers want to see and hear from VMware. Even if they are not implemented, it is good to know that they are aware of community needs.
- Session breakdown of Cluster centered DRS (DRS 1.0)
- Check the clusters state for a need for balancing
- If yes, will select vm to move to another host, and if that has improved the balance metrics then depending on settings for vmotion will migrate the vm
- Session breakdown of vm centered DRS (DRS 2.0)
- Selects a vm
- Finds was described as a “happiness score” for the particular vm on all the hosts in the cluster
- Once the happiness score metric is calculated for each host it will then vmotion to the host which had the highest happiness score
- This will attain vm balance as well as the same drs 1.0 cluster balance at the same time
Sai Inabattini, from the vSphere Performance Group, explained why there is the difference in response times between 1.0 and 2.0 DRS. Current DRS 1.0 runs every 5 minutes takes a snapshot of the cluster and uses that to run the DRS algorithm.
DRS 2.0 has the ability to run every 1 minute due to getting away from the cluster wide snapshot, which happens in memory, which in turn, as he stated, the lighter DRS 2.0 allows for an increase in DRS performance and ability to handle larger inventories.
Other items of serious note are the addition of network as a primary DRS metric, hardware awareness (reliance of vm on underlaying hardware specifics such as virtual GPU), and stable/unstable workload awareness, which will allow DRS to not have constant vmotions within the environment. DRS 2.0 new cost model also tracks power on and power off of vms, which is great depending on if the environment is doing, say, CI/CD (without instant cloning).
Link to the original session > Extreme Performance Series: DRS 2.0 Performance Deep Dive (HBI2880BU)