VM: 1 socket 2 cores OR 2 socket 1 core?

[Afox]

Hello proxmox-community,

I am curious about the question mentioned above. What would be the better setup?

Thanks in advance for every answer

Regards,

Afox

 

[alexadmin]

I too am interested in knowing an answer to this. I usually do 1 socket 2 cores...

 

[swartz]

Hello proxmox-community,

I am curious about the question mentioned above. What would be the better setup?

Thanks in advance for every answer

Regards,

Afox

Assuming the same generation of and same cores in both CPUs, the 1s2c CPU would be marginally faster due to reduced data latency (shorter electrical signal path).
Multi-core CPUs typically share their cache between cores this provides further benefit.

 

[alatteri]

He is talking about CPU topology in guest VMs. Don't think it would make a difference in performance, maybe in licenses of certain softwares though.

 

[Afox]

hello guys,
talking about latency: does proxmox benefit from the improvements below? Their goal is to reduce latency and I am not quite sure if disabling the turbo boost would improve anything..

Table 1. [FONT=&quot]BIOS Settings for Low Latency [/FONT]
[FONT=&quot] [/FONT]
[TABLE="class: MsoNormalTable"]
[TR]
[TD="width: 157"] [FONT=&quot]System Setup Screen [/FONT][FONT=&quot][/FONT]

[/TD]
[TD="width: 157"] [FONT=&quot]Setting [/FONT][FONT=&quot][/FONT]
[/TD]
[TD="width: 157"] [FONT=&quot]Default [/FONT][FONT=&quot][/FONT]

[/TD]
[TD="width: 157"] [FONT=&quot]Recommended Alternative for Low-Latency Environments [/FONT][FONT=&quot][/FONT]

[/TD]
[/TR]
[TR]
[TD="width: 157"] [FONT=&quot]Processor Settings [/FONT]
[/TD]
[TD="width: 157"] [FONT=&quot]Logical Processor [/FONT]

[/TD]
[TD="width: 157"] [FONT=&quot]Enabled [/FONT]
[/TD]
[TD="width: 157"] [FONT=&quot]Disabled [/FONT]
[/TD]
[/TR]
[TR]
[TD="width: 157"] [FONT=&quot]Processor Settings [/FONT]
[/TD]
[TD="width: 157"] [FONT=&quot]Turbo Mode [/FONT]

[/TD]
[TD="width: 157"] [FONT=&quot]Enabled [/FONT]
[/TD]
[TD="width: 157"] [FONT=&quot]Disabled* [/FONT]

[/TD]
[/TR]
[TR]
[TD="width: 157"] [FONT=&quot]Processor Settings [/FONT]
[/TD]
[TD="width: 157"] [FONT=&quot]C-States [/FONT]
[/TD]
[TD="width: 157"] [FONT=&quot]Enabled [/FONT]
[/TD]
[TD="width: 157"] [FONT=&quot]Disabled [/FONT]
[/TD]
[/TR]
[TR]
[TD="width: 157"] [FONT=&quot]Processor Settings [/FONT]
[/TD]
[TD="width: 157"] [FONT=&quot]C1E [/FONT]
[/TD]
[TD="width: 157"] [FONT=&quot]Enabled [/FONT]
[/TD]
[TD="width: 157"] [FONT=&quot]Disabled [/FONT]
[/TD]
[/TR]
[TR]
[TD="width: 157"] [FONT=&quot]Power Management [/FONT]
[/TD]
[TD="width: 157"] [FONT=&quot]Power Management [/FONT]
[/TD]
[TD="width: 157"] [FONT=&quot]Active Power Controller [/FONT]
[/TD]
[TD="width: 157"] [FONT=&quot]Maximum Performance [/FONT]
[/TD]
[/TR]
[/TABLE]

 

[swartz]

He is talking about CPU topology in guest VMs. Don't think it would make a difference in performance, maybe in licenses of certain softwares though.

Oh oops.
In that case for most optimal performance follow whatever design your physical CPU has. Some CPUs have multiple cores sharing single cache (e.g. 2,4,6-core Xeons) in a single NUMA node. While others have multiple groups of NUMA nodes. So it's effectively packaging a 2/3/4-socket design into one physical CPU package with inner-bus(es) connecting NUMA nodes (e.g. high-core count Opterons, Xeons). Each NUMA node with it own cache. Accessing data in local cache is faster than having to traverse NUMA boundary for data in non-local cache.

For optimal performance, vCPU count should not exceed core count in a NUMA node.

So in this case 2 vCores is better than 2 vSockets provided that NUMA and thus CPU affinity is supported (Proxmox 3.4).

Here's a better discussion of NUMA nodes and vCPU affinity.
http://cloudarchitectmusings.com/20...-virtualizing-business-critical-applications/

 

[swartz]

hello guys,
talking about latency: does proxmox benefit from the improvements below? Their goal is to reduce latency and I am not quite sure if disabling the turbo boost would improve anything..

My intuition tells me that Turboboost would reduce latency (as clock speed is increased). However this may be achieved via clocking down other components so as to not overheat the CPU.


Your best bet is to benchmark your application under variety of those settings. CPU design and interaction of various settings is a complex topic and is full of tradeoffs.

 

[Afox]

I checked the settings with or without Turbo Mode and I found out that the latency is reduced by turbo mode. All the named settings brought me a saving of 0,003s (before: min. 0,1785s after: min. 0,1755s). I am not sure about the precentage of savings but I think it is something about 1,7%.

System Setup Screen	Setting	Before	After
Processor Settings	Logical Processor	Enabled	Disabled
Processor Settings	Turbo Mode	Enabled	Enabled
Processor Settings	C-States	Enabled	Disabled
Processor Settings	C1E	Enabled	Disabled
Power Management	Power Management	Maximum Performance	Maximum Performance

 

[Afox]

I am testing further. Referring to the topic I am also interested of how proxmox is sharing vcores. I mean what happens if the number of assigned vcores exceeds the number of "real" vcores? Example: 2 CPUs have 16 vcores (8 each, HT). I assign to different OVZ and KVM 20 Cores.
question a) why is this possible?
question b) does this affect performance?
question c) what if HT is disabled and only 8 "real" cores are available?
question d) Does proxmox environment profit from HT technology?

So many questions. Someone should make a dissertation about this