[SOLVED] [iperf3 speed] Same Node vs 2 Nodes (Found a Bug)

[LnxBil]

Qemu itself doesn't support Properly Numa

I don't know about "properly", yet it does work as it should, as you can see here.

 

[Ramalama]

I don't know about "properly", yet it does work as it should, as you can see here.

We had this discussion already, enabling numa should bind all the "cores" (threads) of a VM to the same Numa-Node.
As long as it's not the case, Numa is almost pointless.

The Actual behaviour still spreads all the Cores of a Single VM to all possible physical cores Randomly without taking Numa-Nodes into Account.

 

[Ramalama]

Just readed again your link:

You mean the second post from ryang? With:

numa0: cpus=0-13,hostnodes=0,memory=30720,policy=bind
numa1: cpus=14-27,hostnodes=1,memory=30720,policy=bind

Thats new lol, i have to try it, thanks for the heading!
If that will work then it will be finally amazing, but even if, it's not documented anywhere on Proxmox and looks like "a guy found something out"...

So in that case, i wouldn't call it "It works as it should", since probably no one knows this and if Proxmox doesn't document it at least (A gui option would be still nicer), its still some sort of an "unknown hack" xD

However Thanks!

EDIT: Mhh, it still doesnt work, i had to reread, its just a definition for the VM-Guest itself, which is pointless again.
You have to use core affinity, which makes things again unmanagable if you have a lot of VM's.
Managing Numa with Core Affinity/Pinning is the only way to make it work, but that we knew already a long time ago. Its not a solution sadly.

Cheers

 

[LnxBil]

You mean the second post from ryang?

Or the 3rd one where I tried it and came to the same conclusion as you did:

Mhh, it still doesnt work, i had to reread, its just a definition for the VM-Guest itself, which is pointless again.
You have to use core affinity, which makes things again unmanagable if you have a lot of VM's.
Managing Numa with Core Affinity/Pinning is the only way to make it work, but that we knew already a long time ago. Its not a solution sadly.

It's a pain to manage by hand, especially with a lot of numa nodes on the hypervisor.