OK, here is a hackish WIP to see if we get anywhere with the L1 concept:
0001 is extracted from a patch from Horiguchi-san to remove the "dead" flag
0002 adds the large bucket, but doesn't use it for anything
0003 uses the new bucket for the L1 cache
0004 changes when to rehash
0005 is Horiguchi-san's v7 benchmark
0006 removes expiration stuff from the benchmark and prevents alloc errors with my patches while running the benchmark
This doesn't align the bucket array to a cacheline boundary, nor does it change the initial number of buckets
make -C contrib/catcachebench/ install
psql -c 'create extension catcachebench'
# This is also from Horiguchi-san
perl
gen_tbl.pl | psql > /dev/null
# warmup
psql -c 'select catcachebench(0)'
# measure median of 3
master:
psql -c 'select catcachebench(1)'
catcachebench
---------------
6084.992169
patch:
./inst/bin/psql -c 'select catcachebench(1)'
catcachebench
---------------
5508.072532
That's decent, but not exactly stellar. I get a huge slowdown in catcachebench(2), however, so I'll have to dig into why before going any further.
Some time I'll also try the function specialization Andres mentioned and see how big of a speedup that gives.
--
John Naylor
EDB:
http://www.enterprisedb.com
