Robert Haas <robertmhaas@gmail.com> writes:
> Nice. What was the overall effect on memory consumption?
Before:
cspell: 31352608 total in 3814 blocks; 37432 free (6 chunks); 31315176 used
After:
cspell: 16214816 total in 1951 blocks; 13744 free (12 chunks); 16201072 used
This is on a 32-bit machine that uses MAXALIGN 8, and I also had
enable_cassert on (hence extra palloc chunk overhead) so it may be
overstating the amount of savings you'd see in production. But it's
in the same ballpark as what Pavel reported originally. AFAICT
practically all of the useful savings comes from the one place he
targeted originally, and the other changes are just marginal gravy.
Before I throw it away, here's some data about the allocations that
go through that code on the Czech dictionary. First column is number
of calls of the given size, second is requested size in bytes:
1 1695 2
1310 3
1965 4
2565 5
1856 6578 7301 8 7 9 2 10
707733 12520 16
107035 20 16 24
22606 28 3 32
8814 36 59 40
4305 44 2 48
2238 52 2 56
1236 60 20 64816 68599 76 1 80408 84 9 88334 92 2 96246 100 1 104164 108 13 112132 116110 124 1 128107 132 3 136
81140 1 144 77 148 40 156 46 164 29 172 39 180 2 184 35 188 31 196 19 204 16 212 12 220 10 228 3 244 1 304 1 400
11120
The spikes at 12/20/28 bytes correspond to SPNodes with 1/2/3 data
items.
BTW, on a 64-bit machine we're really paying through the nose for the
pointers in SPNodeData --- the pointers are bad enough, and their
alignment effects are worse. If we were to try to change this over to
a pointer-free representation, we could probably replace those pointers
by 32-bit offsets, which would save a full factor of 2 on 64-bit.
regards, tom lane