logging

[Hal Murray]

> HPET is a travel out to ACPI system registers mapped into memory, this should
> never be never cached.

Yes.  But there is still the cache for code and data.

This sort of code is amazingly delicate.  Minor changes can make interesting 
changes in the results.

For example:
  for (i = 0; i < BATCHSIZE; i++) {
    clock_gettime(type, &start);  /* warm up cache */
    clock_gettime(type, &stop);
    clock_gettime(type, &start);
    clock_gettime(type, &stop);
...

The extra pair of clock_gettime-s cleaned things up a lot.  At least on one of 
my systems at some point in time.

At one point in time, I got different results when I reran a test.  Sometimes 
it would get duplicate samples.  Try again and they didn't happen.  That was 
on a Pi 1.  Ah.  I just got another example.

      res   avg      min  dups  CLOCK
        1  1069      999 52564  CLOCK_REALTIME
 10000000   820  9999898   -84  CLOCK_REALTIME_COARSE
        1  1075      999 12920  CLOCK_MONOTONIC
        1  1042     1000 62481  CLOCK_MONOTONIC_RAW
        1  1521      999        CLOCK_BOOTTIME

Histogram: CLOCK_REALTIME, 1 ns per bucket, 1000000 samples.
        ns      hits
       999      9577
      1000    930001
      1999       149
      2000      6856
      2999         2
      3000       152
49105 samples were duplicated.
4158 samples were bigger than 3498.


      res   avg      min  dups  CLOCK
        1  1347      999        CLOCK_REALTIME
 10000000  1028  9999890  -109  CLOCK_REALTIME_COARSE
        1  1349      999        CLOCK_MONOTONIC
        1  1279     1000        CLOCK_MONOTONIC_RAW
        1  1838      999        CLOCK_BOOTTIME

Histogram: CLOCK_REALTIME, 1 ns per bucket, 1000000 samples.
        ns      hits
       999      8357
      1000    765081
      1999      4814
      2000    215894
      2999        21
      3000       632
5201 samples were bigger than 3498.

I'm guessing that it's something like the cache conflicts changing because of 
address space randomization.




-- 
These are my opinions.  I hate spam.