Cores compared to Cores + Sockets

[harmonyp]

I am trying to understand the performance differences between creating a VM with 8 cores vs 4 cores/2 sockets. From any tests I manage to do the sysbench results seem to be very similar in fact usually the 4core/2sockets generally does better. While I overload the server (by running both tests at the same time) which is a 8c/16t AMD CPU both also get very similar reduced results.

What would the pros/cons be to using each of the above configs comped to each other.

 

[ermanishchawla]

To understand it better, what is the memory configured for the VM and what is total RAM in server. As this question has more relevance there

 

[harmonyp]

To understand it better, what is the memory configured for the VM and what is total RAM in server. As this question has more relevance there

Both have the same RAM 6GB / 64GB (Node)

 

[Ramalama]

try to enable numa and test again.
on a one socket system it shouldn't matter how you assign the cores.
Without numa it makes, or should make no difference.
Numa on other hand, try to assign the cores corresponding to what can access your ram faster and assign the cores better. To the least used ones, or to the physical cores.

i don't know what numa does exactly in detail, since im simply lazy, but you can read here for example:
https://blog.programster.org/kvm-numa-declaration

All i can say is, it benefits single cpu systems and is needed on multi cpu systems.

cheers

 

[ermanishchawla]

Both have the same RAM 6GB / 64GB (Node)

To test it better set numa flag in cpu and increase memory to more than 50% of total memory to a vm. In that case memory access over upi interconnected will be visible and performance benchmark can be done. Use fio or similar benchmark tool