Обсуждение: Locale timings

I did some "benchmarks" to check whether --enable-locale with LC_ALL=C is
just as fast as --disable-locale, to possibly justify making locale
support the default.  This test only covers locale-aware comparisons,
which seems to be the critical aspect for all intents and purposes.

I loaded a table of a single text column with 454240 rows of English
words.  The table had a size of 21.5 MB.  The values were explicitly
de-sorted, but the order was the same across all test runs.  Then I ran
SELECT * FROM test ORDER BY 1; and timed the wall-clock response time a
few times.  All configuration parameters were left at the default.

The averaged results follow.  Some logarithmic buffering cleverness
appeared to surface, but the results are still distinct enough to be
useful.

no locale:    58s
locale=C:    78s    (ca. 33% slower)
locale=en_US:    118s    (ca. 100% slower)

This confused me, because in my C library a strcoll() call with locale=C
is handed to strcmp() quite directly.  A look into varlena.c:varstr_cmp()
shows that the locale-aware path does some extra copying because there is
no strncoll() function we can use with non-terminated strings.

For testing's sake I replaced the two palloc() calls in that function with
alloca(), which is presumably the fastest possible memory allocator.
Result:

locale=C,alloca:    67s    (ca. 15% slower)

This shows that we're wasting quite a bit of time allocating memory --
probably not only in this place.  I'm pretty sure that the majority of the
rest of the gap comes from the memcpy() operations.  Not that there's a
whole lot we can do about either of these things.

However, I feel that we could reasonably cope with this situation by
replacing

#ifdef USE_LOCALE
/* locale-aware code */
#else
/* non-locale code */
#endif

with

if (locale_is_not_C)
{   /* locale-ware code */
}
else
{   /* non-locale code */
}

This practice should have minuscule impact, and it's probably the plan for
the multibyte side of things as well.

-- 
Peter Eisentraut   peter_e@gmx.net

This is a common way of doing things inside glibc, and the happy result is that 
if you really want to build a non-locale-aware system, you can use a 
compile-time option that replaces the "locale_is_not_C" test with a constant. 
It makes for more maintainable code because there's less chance for bitrot in 
the usual case.

M

Peter Eisentraut wrote:

> I did some "benchmarks" to check whether --enable-locale with LC_ALL=C is
> just as fast as --disable-locale, to possibly justify making locale
> support the default.  This test only covers locale-aware comparisons,
> which seems to be the critical aspect for all intents and purposes.
> 
> I loaded a table of a single text column with 454240 rows of English
> words.  The table had a size of 21.5 MB.  The values were explicitly
> de-sorted, but the order was the same across all test runs.  Then I ran
> SELECT * FROM test ORDER BY 1; and timed the wall-clock response time a
> few times.  All configuration parameters were left at the default.
> 
> The averaged results follow.  Some logarithmic buffering cleverness
> appeared to surface, but the results are still distinct enough to be
> useful.
> 
> no locale:    58s
> locale=C:    78s    (ca. 33% slower)
> locale=en_US:    118s    (ca. 100% slower)
> 
> This confused me, because in my C library a strcoll() call with locale=C
> is handed to strcmp() quite directly.  A look into varlena.c:varstr_cmp()
> shows that the locale-aware path does some extra copying because there is
> no strncoll() function we can use with non-terminated strings.
> 
> For testing's sake I replaced the two palloc() calls in that function with
> alloca(), which is presumably the fastest possible memory allocator.
> Result:
> 
> locale=C,alloca:    67s    (ca. 15% slower)
> 
> This shows that we're wasting quite a bit of time allocating memory --
> probably not only in this place.  I'm pretty sure that the majority of the
> rest of the gap comes from the memcpy() operations.  Not that there's a
> whole lot we can do about either of these things.
> 
> However, I feel that we could reasonably cope with this situation by
> replacing
> 
> #ifdef USE_LOCALE
> /* locale-aware code */
> #else
> /* non-locale code */
> #endif
> 
> with
> 
> if (locale_is_not_C)
> {
>     /* locale-ware code */
> }
> else
> {
>     /* non-locale code */
> }
> 
> This practice should have minuscule impact, and it's probably the plan for
> the multibyte side of things as well.
> 
> 

Peter Eisentraut <peter_e@gmx.net> writes:
> However, I feel that we could reasonably cope with this situation by
> replacing

> #ifdef USE_LOCALE
> /* locale-aware code */
> #else
> /* non-locale code */
> #endif

> with

> if (locale_is_not_C)
> {
>     /* locale-ware code */
> }
> else
> {
>     /* non-locale code */
> }

> This practice should have minuscule impact,

Except perhaps on the size of the executable ;-).  However, USE_LOCALE
affects little enough stuff that that's probably not a big objection.

A possibly more serious objection is whether there are still any systems
out there that don't *have* locale support, and will give us build
errors on <locale.h> or strcoll() or some such.  It looks like
<locale.h> is required by ANSI C, so I think we can get away with this;
does anyone still care about, eg, SunOS 4.1.x?  (One could imagine
providing a stub implementation that equates strcoll() to strcmp() and
so forth, if anyone still does care.)

One nice point is that this will actually make things faster for the
case of locale-enabled-code-running-in-C-locale, which I suspect is not
uncommon, particularly for RPM users.

I'm for it ...
        regards, tom lane