Re: Improve catcache/syscache performance.

Hi,

On 2017-09-13 23:12:07 -0700, Andres Freund wrote:
> Attached is a patch that tries to improve sys/catcache performance,
> going further than the patch referenced earlier.

Here's a variant that cleans up the previous changes a bit, and adds
some further improvements:

Here's the main commit message:

Improve sys/catcache performance.

The following are the individual improvements:
1) Avoidance of FunctionCallInfo based function calls, replaced by
more efficient functions with a native C argument interface.
2) Don't extract columns from a cache entry's tuple whenever matching
entries - instead store them as a Datum array. This also allows to
get rid of having to build dummy tuples for negative & list
entries, and of a hack for dealing with cstring vs. text weirdness.
3) Reorder members of catcache.h struct, so imortant entries are more
likely to be on one cacheline.
4) Allowing the compiler to specialize critical SearchCatCache for a
specific number of attributes allows to unroll loops and avoid
other nkeys dependant initialization.
5) Only initializing the ScanKey when necessary, i.e. catcache misses,
greatly reduces cache unnecessary cpu cache misses.
6) Split of the cache-miss case from the hash lookup, reducing stack
allocations etc in the common case.
7) CatCTup and their corresponding heaptuple are allocated in one
piece.

This results in making cache lookups themselves roughly three times as
fast - full-system benchmarks obviously improve less than that.

I've also evaluated further techniques:
- replace open coded hash with simplehash - the list walk right now
shows up in profiles. Unfortunately it's not easy to do so safely as
an entry's memory location can change at various times, which
doesn't work well with the refcounting and cache invalidation.
- Cacheline-aligning CatCTup entries - helps some with performance,
but the win isn't big and the code for it is ugly, because the
tuples have to be freed as well.
- add more proper functions, rather than macros for
SearchSysCacheCopyN etc., but right now they don't show up in
profiles.

The reason the macro wrapper for syscache.c/h have to be changed,
rather than just catcache, is that doing otherwise would require
exposing the SysCache array to the outside. That might be a good idea
anyway, but it's for another day.

With the attached benchmark for wide tuples and simple queries I get:

pgbench -M prepared -f ~/tmp/pgbench-many-cols.sql

master:
tps = 16112.117859 (excluding connections establishing)
tps = 16192.186504 (excluding connections establishing)
tps = 16091.257399 (excluding connections establishing)

patch:
tps = 18616.116993 (excluding connections establishing)
tps = 18584.036276 (excluding connections establishing)
tps = 18843.246281 (excluding connections establishing)

~17% gain

pgbench -M prepared -f ~/tmp/pgbench-many-cols.sql -c -j 16:
master:
tps = 73277.282455 (excluding connections establishing)
tps = 73078.408303 (excluding connections establishing)
tps = 73432.476550 (excluding connections establishing)

patch:
tps = 89424.043728 (excluding connections establishing)
tps = 89223.731307 (excluding connections establishing)
tps = 87830.665009 (excluding connections establishing)

~21% gain

standard pgbench readonly:
1 client:
master:
tps = 41662.984894 (excluding connections establishing)
tps = 40965.435121 (excluding connections establishing)
tps = 41438.197117 (excluding connections establishing)

patch:
tps = 42657.455818 (excluding connections establishing)
tps = 42834.812173 (excluding connections establishing)
tps = 42784.306987 (excluding connections establishing)

So roughly ~2.3%, much smaller, as expected, because the syscache is
much less of a bottleneck here.

-cj 16:
master:
tps = 204642.558752 (excluding connections establishing)
tps = 205834.493312 (excluding connections establishing)
tps = 207781.943687 (excluding connections establishing)

dev:
tps = 211459.087649 (excluding connections establishing)
tps = 214890.093976 (excluding connections establishing)
tps = 214526.773530 (excluding connections establishing)

So ~3.3%.

I personally find these numbers quite convincing for a fairly localized
microoptimization.

For the attached benchmark, here's the difference in profiles:
before:
single function overhead:
+ 8.10% postgres postgres [.] SearchCatCache
- 7.26% postgres libc-2.24.so [.] __memmove_avx_unaligned_erms
- __memmove_avx_unaligned_erms
+ 59.29% SearchCatCache
+ 23.51% appendBinaryStringInfo
+ 5.56% pgstat_report_activity
+ 4.05% socket_putmessage
+ 2.86% pstrdup
+ 2.65% AllocSetRealloc
+ 0.73% hash_search_with_hash_value
+ 0.68% btrescan
0.67% 0x55c02baea83f
+ 4.97% postgres postgres [.] appendBinaryStringInfo
+ 2.92% postgres postgres [.] ExecBuildProjectionInfo
+ 2.60% postgres libc-2.24.so [.] __strncpy_sse2_unaligned
+ 2.27% postgres postgres [.] hashoid
+ 2.18% postgres postgres [.] fmgr_info
+ 2.02% postgres libc-2.24.so [.] strlen

hierarchical / include child costs:
+ 21.35% 8.86% postgres postgres [.] SearchCatCache

after:
single function overhead:
+ 6.34% postgres postgres [.] appendBinaryStringInfo
+ 5.12% postgres postgres [.] SearchCatCache1
- 4.44% postgres libc-2.24.so [.] __memmove_avx_unaligned_erms
- __memmove_avx_unaligned_erms
+ 60.08% appendBinaryStringInfo
+ 13.88% AllocSetRealloc
+ 11.58% socket_putmessage
+ 6.54% pstrdup
+ 4.67% pgstat_report_activity
+ 1.20% pq_getbytes
+ 1.03% btrescan
1.03% 0x560d35168dab
+ 4.02% postgres postgres [.] fmgr_info
+ 3.18% postgres postgres [.] ExecBuildProjectionInfo
+ 2.43% postgres libc-2.24.so [.] strlen

hierarchical / include child costs:
+ 6.63% 5.12% postgres postgres [.] SearchCatCache1
+ 0.49% 0.49% postgres postgres [.] SearchSysCache1
+ 0.10% 0.10% postgres postgres [.] SearchCatCache3

(Most of the other top entries here are addressed in neirby threads)

- Andres