Povray on Quad Xeon E5-4624L and Quad Opteron 6344

[jhu]

40+ threads of rendering goodness. I'm surprised at the performance degradation with hyperthreading on under Windows. Makes no sense, even under full hardware threads.

 

[Burpo]

Windows Server 2008 R2? :\

 

[jhu]

Windows Server 2008 R2? :\

Server version of Windows 7.

 

[monstercameron]

I'm not quite understanding what this is testing, can anyone break it down or distill it?

 

[AtenRa]

@jhu thx nice one

@monstercameron Per core scaling in PovRay

 

[jhu]

I'm not quite understanding what this is testing, can anyone break it down or distill it?

Performance scaling based on number of threads and cores. That quad Xeon has a total of 40 cores and 80 hardware threads. Ideally, with 40 threads, there should be a 40x speedup in render times compared with one core. As you can see, the max is only about 24x to 29x. With HT on under Windows, it's even less.

 

[monstercameron]

Performance scaling based on number of threads and cores. That quad Xeon has a total of 40 cores and 80 hardware threads. Ideally, with 40 threads, there should be a 40x speedup in render times compared with one core. As you can see, the max is only about 24x to 29x. With HT on under Windows, it's even less.

ok and why do you think that is? and why doesnt the piledriver opteron have similar results?

 

[jhu]

ok and why do you think that is? and why doesnt the piledriver opteron have similar results?

1) I don't know, but I would guess memory access congestion due to so many threads.

2) The Opterons do have similar results: it's not 1:1 performance increase at high core counts.

 

[Sequences]

They recompiled povray on linux but didn't do so on Windows. I'm not sure how much that would make a difference, but something to keep in mind.

 

[Ajay]

They recompiled povray on linux but didn't do so on Windows. I'm not sure how much that would make a difference, but something to keep in mind.

And they use VS10 instead of Intel Composer XE, etc...

----------------------------------------------------------------

The drop off w/HT could be as simple as locking a thread to a logical core, that not how HT is designed to work.

I suppose, in reality, none of that really matters to end users who aren't using povray as part of a rendering system where they have their own developers optimizing the executables.

 

[Nothingness]

1) I don't know, but I would guess memory access congestion due to so many threads.

Unlikely given that there's a table that shows that under Linux HT brings some small speedups (with 40 cores, 40s, with 40+HT, 32s), and on top of that it's ~20% faster so memory congestion would be even more of an issue.