What's the best way to demonstrate the effect of affinity setting?

Once I noticed that Windows doesn't keep computation-intensive threads on a specific core - it keeps switching cores instead. So I speculated that the job would be done faster, if the thread would keep access to the same data caches. And really, I was able to observe a stable ~1% speed improvement after setting the thread's affinity mask to a single core (in a ppmd (de)compression thread). But then I tried to build a simple demo for this effect, and more or less failed - that is, it works as expected on my system (Q9450):

buflog=21 bufsize=2097152
(cache flush) first run    = 6.938s
time with default affinity = 6.782s
ti开发者_JAVA技巧me with first core only  = 6.578s
speed gain is 3.01%

but people I asked weren't exactly able to reproduce the effect. Any suggestions?

#include <stdio.h>
#include <windows.h>
int buflog=21, bufsize, bufmask;
char* a;
char* b;
volatile int r = 0;
__declspec(noinline)
int benchmark( char* a ) {
  int t0 = GetTickCount();
  int i,h=1,s=0;
  for( i=0; i<1000000000; i++ ) {
    h = h*200002979 + 1;
    s += ((int&)a[h&bufmask]) + ((int&)a[h&(bufmask>>2)]) + ((int&)a[h&(bufmask>>4)]);
  } r = s;
  t0 = GetTickCount() - t0;
  return t0;
}
DWORD WINAPI loadcore( LPVOID ) {
  SetThreadAffinityMask( GetCurrentThread(), 2 );
  while(1) benchmark(b);
}
int main( int argc, char** argv ) {
  if( (argc>1) && (atoi(argv[1])>16) ) buflog=atoi(argv[1]);
  bufsize=1<<buflog; bufmask=bufsize-1;
  a = new char[bufsize+4];
  b = new char[bufsize+4];
  printf( "buflog=%i bufsize=%i\n", buflog, bufsize );
  CreateThread( 0, 0, &loadcore, 0, 0, 0 );
  printf( "(cache flush) first run    = %.3fs\n", float(benchmark(a))/1000 );
  float t1 = benchmark(a); t1/=1000;
  printf( "time with default affinity = %.3fs\n", t1 );
  SetThreadAffinityMask( GetCurrentThread(), 1 );
  float t2 = benchmark(a); t2/=1000;
  printf( "time with first core only  = %.3fs\n", t2 );
  printf( "speed gain is %4.2f%%\n", (t1-t2)*100/t1 );
  return 0;
}

P.S. I can post a link to compiled version if anybody needs that.

---

default affinity:

_{(source: dreamhosters.com)}

affinity set to core #4

_{(source: dreamhosters.com)}

Now, this is an archiver. Do you really think that the worker thread going all around the cpu is ok?

---

Maybe you are just lucky, and on the other PCs where you tested the program, someone did exactly the same thing as you did, but his thread is sleeping a lot.

That would lead to your program being interrupted every now and then, when the other thread gets scheduled.

---

How do you know the other 3 cores are being used by your thread and not some system threads? For example if you are paging or something. Set up some performance counters on your process in perfmon and verify this assumption.

---

1. Windows doesn't deliberately swap processes between CPUs. If it did it to you, you were just unlucky.
2. You might get minor speed breaks if you are getting a lot of cache hits, it depends on your application. (Unless you have some big iron with funky NUMA memory architecture, that can cause all sorts of dependencies).
3. In your case, why not just increase the process priority so that it never gets swapped off the CPU?