Evaluating host available memory when starting virtual machines

[havengit]

problem

I'm not sure if this is a bug.
The situation is as follows:
When starting a virtual machine or performing a live migration, ACS evaluates whether the target host has sufficient memory. As shown in the screenshot, the total memory is 93GB, with 86GB currently used and 71GB allocated. Attempting to start an 8GB virtual machine succeeds. Using free -g confirms the system has indeed used 86GB.
Analysis of the logs indicates that ACS determines whether the host has sufficient memory resources to start a VM based on allocated memory, not actual used memory.

The system and other programs consume a certain amount of memory. Typically, the amount of memory used is higher than the amount allocated, which can lead to insufficient memory and trigger an OOM (Out of Memory) issue.

I know that setting host.reserved.mem.mb in the agent configuration file reserves memory. This is a static configuration. I'm unsure if there is a better approach.

Translated with DeepL.com (free version)

VM Start log.txt

versions

Cloud stack : 4.18.0.0
Hypervisor: KVM 6.2/ Libvirt 8.0
OS: Ubuntu 22.04 LTS

The steps to reproduce the bug

The steps have been described above.

What to do about it?

I'm not sure if this is a bug or if it's just by design. Is there a better way to avoid this issue?

[prashanthr2]

@havengit This behavior is actually by design. CloudStack looks at allocated capacity rather than used to ensure a VM always has access to its full RAM when it needs it. For example, if VM1 is allocated 8GB but is currently only using 1GB, and CloudStack were to ignore that allocation and assign the 'unused' 7GB to VM2, you’d run into a major conflict the moment VM1 actually tries to use its full 8GB.

If you're concerned about OOM (Out of Memory) issues, there are a few things you can do:

1. Make sure your mem.overprovisioning.factor is set to 1.0. This ensures you aren't 'selling' more RAM than the hardware physically has.

2. As you mentioned, setting a value for host.reserved.mem.mb is a great move. It carves out a dedicated slice of memory specifically for the host OS and agent processes.

3. You can also use cluster.memory.allocated.capacity.disablethreshold to cap things. For example, setting it to 0.9 tells CloudStack to stop deploying new VMs once the cluster hits 90% allocation. Just keep in mind that this specific setting is applied at the cluster level rather than per-host.

[havengit]

Thank you for your reply. I now understand the process of allocating memory. I have no further questions.