AMD Opteron 6276 - Slow performance compared to bare metal or ESXi

[EpicLPer]

Heya!

Long story short I'm currently looking into Proxmox since ESXi has some pretty weird crash-happy issues with one of my eSATA cards.
Since it's my first time using Proxmox I tried configuring and looking around on how to properly optimize VMs, then set up a Win 2022 Server and gave it 64 Cores + 32GB RAM. I compared some Benchmark results, namely Cinebench right now, to bare metal vs ESXi vs Proxmox, and so far the Proxmox Windows VM only has a score of barely ~13000 while bare metal and ESXi both have ~21000.

I've tried various different combinations, such as enabling/disabling NUMA, enabling/disabling RAM ballooning, changing from SATA to SCSi on the main SSD, changing graphics etc. etc. etc., but every single config combo so far resulted in just about the same above mentioned Cinebench score of around 13000 which is way too low for that hardware.

Is there anything I'm missing still? I'm probably for sure doing something wrong since it's my first time with Proxmox after all ^^
Thanks already in advance!

 

[czechsys]

Opteron 6276 is very old and slow performant. Use better HW. Problem solved.

 

[EpicLPer]

Opteron 6276 is very old and slow performant. Use better HW. Problem solved.

That's not really a proper answer in regard to bare metal or ESXi performing almost twice as fast with similar settings and same hardware.

 

[B.Otto]

Hey,

a quick search yields that the Opteron 6276 has 16C/16T. How many CPUs (sockets) do your server have? A general rule is that a single VM should not have more vCPUs than physically availiable. (it is okay though that the total number of vCPUs - spread over several VMs - can be higher than the total number of CPUs).

I also dislike giving multiple sockets to one VM, though you (rightly) did enable NUMA.

Other than that I dont really have an idea what causes the difference in performance...

Kind regards,
Benedikt

 

[EpicLPer]

Hey,

a quick search yields that the Opteron 6276 has 16C/16T. How many CPUs (sockets) do your server have? A general rule is that a single VM should not have more vCPUs than physically availiable. (it is okay though that the total number of vCPUs - spread over several VMs - can be higher than the total number of CPUs).

I also dislike giving multiple sockets to one VM, though you (rightly) did enable NUMA.

Other than that I dont really have an idea what causes the difference in performance...

Kind regards,
Benedikt

Yeah, I've tried with both 1 Socket and 4 Sockets in the Hardware tab, and the server indeed has combined 64 Cores spread over 4 CPUs.

 

[LnxBil]

Can you see the full utilization of cinebench in your VM AND on your PVE host?

 

[EpicLPer]

Can you see the full utilization of cinebench in your VM AND on your PVE host?

Yeah, and it does use 100% host CPU when the benchmark is running and only ever so often dips to 30% CPU usage, but that is most likely normal. Only happens very occasionally.