Re: GiST index performance

On Thu, 11 Jun 2009, Tom Lane wrote:
> So it'd be worth converting your functions to V1 style.

Does that produce a significant reduction in overhead? (You'll probably
say "yes, that's the whole point").

> hmm ... memcpy or qsort maybe?

Surprise:

CPU: Core 2, speed 1998 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Clock cycles when not halted) with a unit mask of 0x00 (Unhalted core cycles) count
100000
samples  %        image name               app name                 symbol name
3005354  40.2868  libc-2.9.so              libc-2.9.so              __mcount_internal
1195592  16.0269  libc-2.9.so              libc-2.9.so              mcount
549998    7.3727  postgres                 postgres                 gistnext
420465    5.6363  postgres                 postgres                 fmgr_oldstyle
376333    5.0447  no-vmlinux               no-vmlinux               (no symbols)
210984    2.8282  postgres                 postgres                 FunctionCall5
182509    2.4465  postgres                 postgres                 gistdentryinit
174356    2.3372  btree_gist.so            btree_gist.so            gbt_int4_consistent
142829    1.9146  postgres                 postgres                 FunctionCall1
129800    1.7400  postgres                 postgres                 .plt
119180    1.5976  nvidia_drv.so            nvidia_drv.so            (no symbols)
96351     1.2916  libxul.so                libxul.so                (no symbols)
91726     1.2296  btree_gist.so            btree_gist.so            gbt_num_consistent

A quick grep in the postgres source for mcount reveals no hits. No idea
what it does - there is no man page for it.

Matthew

--
 I pause for breath to allow you to get over your shock that I really did cover
 all that in only five minutes...        -- Computer Science Lecturer