* Re: mdocml: Gprof(1) is fun.
[not found] ` <20140106235620.GA31237@britannica.bec.de>
@ 2014-01-07 0:22 ` Ingo Schwarze
0 siblings, 0 replies; only message in thread
From: Ingo Schwarze @ 2014-01-07 0:22 UTC (permalink / raw)
To: tech
Hi,
Joerg Sonnenberger wrote on Tue, Jan 07, 2014 at 12:56:20AM +0100:
> On Mon, Jan 06, 2014 at 06:46:07PM -0500, schwarze@mdocml.bsd.lv wrote:
>> Do not copy predefined strings into the dynamic string table, just
>> leave them in their own static table and use that one as a fallback
>> at lookup time. This saves us copying and deleting them for each manual.
>> No functional change.
> What about doing a binary sort of that table and also storing the size
> and checking that first?
Nice idea, that might help a bit more.
There is one other trivial improvement - caching the result of
the uname(3) multi-sysctl(3) library call, which is likely to
yield another 4-5% speedup, and after that, the gprof(1) analysis
probably ought to be redone to find the areas that are now taking
the biggest chunks of time. Quite possible what you say may be
among them.
Here is a krautcomputing backup of my first gprof(1) analysis,
mostly such that i don't lose it...
Yours,
Ingo
----- 8< ----- schnipp ----- >8 ----- 8< ----- schnapp ----- >8 -----
All numbers in percent of total user+system time of the profiled
mandocdb(8) executable when run with -Q over /usr/share/man/.
Profiling overhead (55% total):
-------------------------------
__mcount 35.3
write 8.5 43.8
fsync 5.8 49.6
_thread_sys_unlink 2.6 52.2
_thread_sys_lseek 2.5 54.7
fcntl 0.8 55.5
In the following call tree,
* "->" marks functions where different code paths join,
coming down from different directions. This is only noted
where more than one of these paths is relevant for performance.
* "!!!" marks standard library functions where a lot of time
was wasted in a particular code path. These were marked
as a first step to identify candidates for optimization.
* "<<<" marks mandoc(3) functions selected for optimization,
as a second step in deciding what to work on.
* "DONE" refers to code optimized on Jan 6, 2014;
all these optimizations are committed.
The total speedup archieved so far is 45% (3.1 -> 1.7 seconds),
which is nearly a factor of two.
Time spent on payload (45% total, about three seconds absolute)
---------------------------------------------------------------
main mandocdb 44.6
mpages_merge 42.9
mparse_readfd 33.5
mparse_parse_buffer 30.5
mparse_buf_r 30.0
roff_parseln 17.3
roff_Dd 10.5 DONE
roff_setstr 17.6 DONE
roff_setstrn 17.5 DONE <<< TARGET 1 <<<
strcmp 5.1 DONE !!!
mandoc_malloc 4.1 DONE
malloc 4.0 DONE !!!
mandoc_strndup 3.0 DONE
mandoc_malloc DONE -> roff_setstrn
memcpy 1.4 DONE !!!
roff_parse 2.9
roff_getstrn 2.3 <<< TARGET 2 <<<
strncmp 1.4 !!!
roff_TH 1.6 DONE
roff_setstr DONE -> roff_Dd
roff_cond_sub 1.5
roff_parse -> roff_parseln
roff_ds 0.8
roff_setstr -> roff_Dd
mdoc_parseln 7.9
mdoc_pmacro 7.7
mdoc_macro 6.3
blk_full 3.9
rew_sub 1.7
rew_last 3.6
mdoc_valid_post 5.7
post_dd 4.1 DONE
mandoc_normdate 5.7 DONE <<< TARGET 3 <<<
a2time 5.0 DONE
mktime 4.9 DONE !!!
time2a 0.8 DONE
strftime 0.5 DONE !!!
post_os 1.0
uname 1.0 !!!
dword -> in_line_eoln
in_line_eoln 2.3
dword 2.3
mdoc_word_alloc 1.9
node_append 2.4
mdoc_valid_post
roff_strdup 1.1
mandoc_realloc 0.8
realloc 0.7
rew_elem 1.6
rew_last -> rew_sub
in_line
man_parseln 1.8
man_pmacro 1.7
blk_imp 1.1
rew_scope 0.7
man_unscope 0.9
man_valid_post 1.7
post_TH 1.7 DONE
mandoc_normdate DONE -> post_dd
in_line_eoln 0.6
mparse_end 0.5
_thread_sys_open 1.8
_thread_sys_close 1.0
munmap 0.8
mparse_reset 8.8
roff_reset 8.6 DONE
roff_setstr DONE ->roff_Dd
roff_free1 4.2 DONE
roff_freestr 4.0 DONE <<< TARGET 4 <<<
free 4.6 DONE !!!
munmap DONE -> mparse_readfd
treescan 1.6
fts_read 1.4
--
To unsubscribe send an email to tech+unsubscribe@mdocml.bsd.lv
^ permalink raw reply [flat|nested] only message in thread