Question about Proxmox Cluster Resource Scheduling

[josephnp]

Proxmox Cluster Resource Scheduling currently running with these settings

Scheduling Mode - Dynamic Load
Rebalance on Start - Check
Automatic Rebalance - Check
Imbalance Threshold - 30
Rebalancing Method = TOPSIS
Hold duration - default (3)
Minimum imbalance improvement = 10

There is a VM that's usually using 90 - 95% of its assigned memory of 128 GB. It's constantly being migrated between the nodes and I'm not sure why. I thought the migration only happens if the node is being pressurized, instead of just the VM. The nodes in the cluster have 768 GB of memory and don't have much load on it. The VM workload profile is memory intensive.

 

[Francois Baillargeon]

I'd try and raise the minimum imbalance to say 25 so it stops doing it so frequently see where it lands.

You could also change the hold duration to a higher value

 

[josephnp]

Regardless of threshold/margin/hold-duration tuning (within reasonable ranges), the reported cluster imbalance sits around 36-40%, and the 128GB/95%-utilized VM gets migrated repeatedly without the imbalance meaningfully or durably dropping afterward. I had to up the threshold to 45% to stop it.

Here's my cluster setup at the moment:

5 nodes, each with 768 GB memory
1 HA-managed VM per node, sizes ranging from 32 GB to 256 GB
One VM (128 GB allocated) consistently runs at ~95% of its own assigned memory; the other 4 VMs use a much smaller fraction of their respective allocations
Overall node-level resource usage is low across the board — nowhere near capacity on any node

From the docs, cluster imbalance is derived from the standard deviation and mean of individual node loads, normalized to 0-100%. Given that I only have 1 VM per node, and one VM is using a much higher fraction of its allocation than the others, my working theory is that this naturally produces a persistently high relative spread (something like a coefficient of variation) even though absolute utilization is low cluster-wide — and that moving the VM around doesn't fix this because the outlier just changes nodes rather than the underlying variance improving.

Given this topology, is there a recommended way to configure CRS so it doesn't chase a statistically unfixable imbalance — e.g., is excluding a single VM from dynamic rebalancing (while keeping it in HA) supported, or is a much higher static threshold the only practical option?

 

[UdoB]

e.g., is excluding a single VM from dynamic rebalancing (while keeping it in HA) supported

Sure: in <your vm> - top right "More" - "Manage HA" you can find two checkboxes, one being "Auto-Rebalance [x]".

I did not test the actual behavior, but I would give it a try...