Testing of branch windows_numa

[andygrant]

Testing of branch windows_numa

[AndyGrant]

Resolved, soon to be merged.

Thanks for your help :)

[CoffeeOne]

Change:

-  fun2_t fun2 = (fun2_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");
-  fun3_t fun3 = (fun3_t)(void(*)())GetProcAddress(k32, "SetThreadGroupAffinity");
+  fun2_t fun2 = (fun2_t)GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");
+  fun3_t fun3 = (fun3_t)GetProcAddress(k32, "SetThreadGroupAffinity");

$ make
gcc -DNDEBUG -O3 -std=gnu11 -Wall -Wextra -Wshadow -march=native -flto *.c fathom/tbprobe.c -lpthread -lm -DUSE_POPCNT -msse3 -mpopcnt -o Ethereal
windows.c: In function 'bestGroup':
windows.c:39:25: warning: cast between incompatible function types from 'FARPROC' {aka 'long long int (*)()'} to 'void (*)()' [-Wcast-function-type]
   fun1_t fun1 = (fun1_t)(void(*)())GetProcAddress(k32, "GetLogicalProcessorInformationEx");
                         ^
windows.c:39:17: warning: cast between incompatible function types from 'void (*)()' to 'int (*)(LOGICAL_PROCESSOR_RELATIONSHIP,  struct _SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *, DWORD *)' {aka 'int (*)(enum _LOGICAL_PROCESSOR_RELATIONSHIP,  struct _SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *, long unsigned int *)'} [-Wcast-function-type]
   fun1_t fun1 = (fun1_t)(void(*)())GetProcAddress(k32, "GetLogicalProcessorInformationEx");
                 ^
windows.c: In function 'bindThisThread':
windows.c:103:17: warning: cast between incompatible function types from 'FARPROC' {aka 'long long int (*)()'} to 'int (*)(USHORT,  struct _GROUP_AFFINITY *)' {aka 'int (*)(short unsigned int,  struct _GROUP_AFFINITY *)'} [-Wcast-function-type]
   fun2_t fun2 = (fun2_t)GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");
                 ^
windows.c:104:17: warning: cast between incompatible function types from 'FARPROC' {aka 'long long int (*)()'} to 'int (*)(void *, const GROUP_AFFINITY *, struct _GROUP_AFFINITY *)' {aka 'int (*)(void *, const struct _GROUP_AFFINITY *, struct _GROUP_AFFINITY *)'} [-Wcast-function-type]
   fun3_t fun3 = (fun3_t)GetProcAddress(k32, "SetThreadGroupAffinity");
```                 ^

[CoffeeOne]

@AndyGrant One more comment about that.
stockfish has this code:

void Thread::idle_loop() {

  // If OS already scheduled us on a different group than 0 then don't overwrite
  // the choice, eventually we are one of many one-threaded processes running on
  // some Windows NUMA hardware, for instance in fishtest. To make it simple,
  // just check if running threads are below a threshold, in this case all this
  // NUMA machinery is not needed.
  if (Options["Threads"] >= 8)
      WinProcGroup::bindThisThread(idx);

Do you want to do something similar in Ethereal?
I connected a 64 core machine under Windows with http://chess.grantnet.us
and had super weird effects (only 16% cpu load with 60 threads) and only 0.1 MN/s.
I am quite sure it is because of that. Most likely all the threads were bound to the same physical node.

[AndyGrant]

Okay, this is easy enough to do.
What about the case when we run tests using 8threads at once with your machine?
It looks to me like this solution only pushes back the problem.

[CoffeeOne]

You are right, the problem comes again at 8 threads with stockfish. But the vast majority of tests are 1 thread tests.

[AndyGrant]

Okay, well do you have any aversion to using Threads > 8 instead of Threads >= 8?

Stockfish usually does 7thread tests, but I do 8thread since the majority of workers are connected on 8 threads.

If not, I will submit a dummy patch where both master/test have the numa change

[CoffeeOne]

No aversion :), I always use the full number of cores/threads when I use the engine for analysis or for engine-engine games in the internet ;)

[AndyGrant]

dd74700 resolved