[Linux-ia64] strange cache effect

[Linux-ia64] strange cache effect

[Frederic Wagner]

Hi everyone,

I have an account on an itanium quadriprocessor running
linux 2.4.18 and gcc 2.96

when compiling the following code with gcc -O2 -Os

#define N 20000000
int B[N + 25000] __attribute__ ((__aligned__ (16384)));

int
init (void)
{
  int i, n, pgsz = getpagesize ();

  n = pgsz / sizeof (int);
  printf ("n=%d pgsz=%d B=%p sizeof=%lu\n", n, pgsz, B, sizeof (B));

  for (i = 0; i < N + 20480; i += n)
    B[i] = 0;
  return 0;
}

int
doit (void)
{
  int i, j, x = 0;
  for (i = 0; i < N; i++)
    for (j = 0; j < 100; j += 8)
      x += B[j] + B[j + 1024] + B[j + 2048] + B[j + 3072] + B[j + 4096];

  return x;
}

int
main (int argc, char **argv)
{
  init ();
  printf ("%d\n", doit ());
  return 0;
}

I get a 
        addl r14 = @ltoff(B#), gp
	        ;;
        ld8 r14 = [r14]

in the main loop nest (I verified by disassembling)

the problem is the following:
when alone on the machine, with no load
the ld8 r14 = [r14] generates either 1 miss every time or 0 miss
(in L1 data)

ie:

[clauss@sigmicroia64 last]$ time pfmon
-eL1D_READ_MISSES_RETIRED,LOADS_RETIRED
--drange=0x6000000000000cb0-0x6000000000000db0 ./code
n@96 pgsz\x16384 B=0x6000000000008000 sizeof€100000
0
            260002762 L1D_READ_MISSES_RETIRED
	    260005474 LOADS_RETIRED

real    0m8.793s
user    0m8.686s
sys     0m0.108s
[clauss@sigmicroia64 last]$ time pfmon
-eL1D_READ_MISSES_RETIRED,LOADS_RETIRED
--drange=0x6000000000000cb0-0x6000000000000db0
./code
n@96 pgsz\x16384 B=0x6000000000008000
sizeof€100000
0
                329 L1D_READ_MISSES_RETIRED
          260005474 LOADS_RETIRED

real    0m7.494s
user    0m7.383s
sys     0m0.112s


note that --drange is used to monitor only the wanted load
no recompilation, I just run it several times (more than 100 times
now I think :), and it just oscillates between these two values.

It can't be a measure problem as the time increases too

so first question: why that ?

what makes me think it may be related to system is that
when increasing the system load (i start 4 other processes)
it stabilizes .... but on the best case !
that is I have always values like
                 3442 L1D_READ_MISSES_RETIRED
            260005474 LOADS_RETIRED

with small variations which are due to load 
but it doesn't get close to number of loads any more

As I don't want to break Amdahl's law if anyone here has a 
suggestion it would be welcome ;-)

thanks for your attention

Wagner Fred

-- 
-----------------------------------------------------------------------
Unix - where you can throw the manual on the keyboard and get a command