Leela much faster on linux?

[Cyan]

Didn't expect there to be such a huge difference, what could be causing this?

Code: Select all

10000 evaluations: Ubuntu > Windows > OS X
                   4608   > 2155    > 980  p/s

CPU：intel i7 7700
GPU：nvidia GTX1070

Windows 10:
leela: benchmark
200000 games in 7.59 seconds -> 26350 g/s (3293 g/s per thread)
Avg Len: 451.53 Score: -1.989495

Leela: netbench
2000 predictions in 2.03 seconds -> 985 p/s
10000 evaluations in 4.64 seconds -> 2155 p/s

OS X 10.12.5：
Leela: benchmark
200000 games in 6.63 seconds -> 30165 g/s (3770 g/s per thread)
Avg Len: 451.58 Score: -1.502100

Leela: netbench
2000 predictions in 3.76 seconds -> 531 p/s
10000 evaluations in 10.20 seconds -> 980 p/s

Ubuntu 16.04.2：
Leela: benchmark
200000 games in 6.05 seconds -> 33057 g/s (4132 g/s per thread)
Avg Len: 451.72 Score: -1.756510

Leela: netbench
2000 predictions in 1.68 seconds -> 1190 p/s
10000 evaluations in 2.20 seconds -> 4545 p/s

Ubuntu 17.04:
Leela: benchmark
200000 games in 6.40 seconds -> 31250 g/s (3906 g/s per thread)
Avg Len: 451.80 Score: -1.679700

Leela: netbench
2000 predictions in 1.65 seconds -> 1212 p/s
10000 evaluations in 2.17 seconds -> 4608 p/s

Source: https://tieba.baidu.com/p/5122485560

[Drew]

Assuming code base is equivalent, Windows 10/7/whatever has much more system overhead than Ubuntu. That alone will yield a performance increase because more system resources are available for Leela to use on Linux.

[Mike Novack]

Not really an explanation for a significant difference when the programs we are dealing with are "crunches". The relative efficiencies of the operating systems should matter only when a significant percentage of the time is spent in them (system calls, other dynamic calls).

Maybe these programs need "tuning"? << example, a constantly used dynamic call converted to pseudo static >> This sort of thing used to be my line of country in my working days, finding unfortunate choices made by the programmers and replacing with something that might be an order of magnitude faster but still easy to maintain. << those were nice assignments when I got them because often got awarded a bonus of a percent or two of the saving in annual CPU charges >>

[Garf]

The relative efficiencies of the operating systems should matter only when a significant percentage of the time is spent in them (system calls, other dynamic calls).

He's testing the GPU acceleration, so that's exactly what happens. The CPU is forwarding the work to the GPU, through the OS, graphics drivers, OpenCL stack, etc.