The problem is that you are using the numeric values of the hash in a string context: print $fh "$key $hash{$key}\n";; thus perl converts what were IVs (internal integers) into PVs (internal strings), caching the result in the expectation you might use them in a string context again.

As an IV, each value requires 24 bytes, but once converted to a string and stored in a PV, each value requires ~56 bytes.

You can avoid the conversion being cached by making perl convert a temporary value to a string for printing like this:

     printf $fh "$key %d\n", 0+$hash{$key};
[download]

That will avoid the memory growth.

You store a number then access it as a string so Perl caches the stringised version of the number along with the number. Try quoting the value so it's stored as a string instead of as a number to check (works for me - size is stable).

I am not experienced enough to know most of the inner details, but it's fair to say that perl does a lot of work internally when the items are first accessed. If you change to Devel::Peek, you can see how the structure changes after access:

use strict;
use warnings;
use Devel::Peek qw(Dump);

my $s = 'AAAAAAAAAAAAAAA';

my %hash = map {$s++ => 1} 1 .. 2;

print "before:\n";
Dump \%hash;


open my $fh, '>', 'j1.txt' or die $!;

for my $key (keys %hash) {
    print $fh "$key $hash{$key}\n";
}

print "after:\n";

Dump \%hash;
[download]

Output:

before:
SV = IV(0x1e5cca8) at 0x1e5ccb8
  REFCNT = 1
  FLAGS = (TEMP,ROK)
  RV = 0x1e88420
  SV = PVHV(0x1e63ef0) at 0x1e88420
    REFCNT = 2
    FLAGS = (PADMY,SHAREKEYS)
    ARRAY = 0x1ed7880  (0:6, 1:2)
    hash quality = 125.0%
    KEYS = 2
    FILL = 2
    MAX = 7
    Elt "AAAAAAAAAAAAAAA" HASH = 0xa39c9065
    SV = IV(0x1e7a9a8) at 0x1e7a9b8
      REFCNT = 1
      FLAGS = (IOK,pIOK)
      IV = 1
    Elt "AAAAAAAAAAAAAAB" HASH = 0x7747aa6e
    SV = IV(0x1e7a990) at 0x1e7a9a0
      REFCNT = 1
      FLAGS = (IOK,pIOK)
      IV = 1

after:
SV = IV(0x1e883e0) at 0x1e883f0
  REFCNT = 1
  FLAGS = (TEMP,ROK)
  RV = 0x1e88420
  SV = PVHV(0x1e63ef0) at 0x1e88420
    REFCNT = 2
    FLAGS = (PADMY,OOK,SHAREKEYS)
    ARRAY = 0x1e75110  (0:6, 1:2)
    hash quality = 125.0%
    KEYS = 2
    FILL = 2
    MAX = 7
    RITER = -1
    EITER = 0x0
    RAND = 0x69863a90
    Elt "AAAAAAAAAAAAAAA" HASH = 0xa39c9065
    SV = PVIV(0x1e7f860) at 0x1e7a9b8
      REFCNT = 1
      FLAGS = (IOK,POK,pIOK,pPOK)
      IV = 1
      PV = 0x1ed7810 "1"\0
      CUR = 1
      LEN = 16
    Elt "AAAAAAAAAAAAAAB" HASH = 0x7747aa6e
    SV = PVIV(0x1e7f878) at 0x1e7a9a0
      REFCNT = 1
      FLAGS = (IOK,POK,pIOK,pPOK)
      IV = 1
      PV = 0x1e98b90 "1"\0
      CUR = 1
      LEN = 16
[download]

Here's a brief on memory usage.

It's not the "first access" that does it, it's the "give it to me different". Stored as a number and accessed as a string in the OP's case.

Premature optimization is the root of all job security

stevieb

Thanks for the nice link to Mastering Perl. It was helpful and I learned from all the replies given here. In a thousand years, I don't think I would have caught the distinction between using the hash value as a string vs. a number.