Unable to achieve desired speed up using parfor

Hi,

I am initializing several instances of a matlab (p)code using parfor loops on two computers with the following configurations.

Comp A: 16 core 3.4GHz, 8GB per core @ 3200MHz,

Comp B: 32 core 3.6GHz, 8GB per core @ 3200MHz,

I am launching 16 instances on A and 32 on B. I find that all instances on B finish in about half the time as those on A. It baffles me since the spec scale almost identically. Also, all instances do the same thing, hence identical computational overhead. Is there any hardware optimization that should be done for better efficiency on A?

[Walter Roberson]

Question: Do the systems have 16 physical cores and 32 physical cores, or do they have 16 and 32 logical cores?

[Balachandra Suri]

Physical cores. Ryzens 5950x vs 5975wx

[埃博拉酱]

You've provided so little information as to what might be causing this that I don't even know where to start.

[Balachandra Suri]

Please let me know what other information can be useful? Motherboard config?

[Rik]

My initial guess was that the generation would be different and hence the number of instructions per cycle may be different. That doesn't seem to be the case here.

Perhaps it is the cache? If everything fits in the CPU cache there is no need to go to RAM. I don't have any other plausible cause, unless the smaller chip doesn't actually reach the frequency you mentioned due to thermal and/or power throttling.

Answer 1

[Matt J]

I find that all instances on B finish in about half the time as those on A. It baffles me...

That is the expected result, assuming you are running the same loop on both computers. Assuming for example that it is a 32 iteration loop,

parfor i=1:32
    ...
end

then Comp A would be assigned 2 iterations per core, while Comp B will be assigned only 1. So, it makes perfect sense that Comp B will finish in half the time.