| From: | Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> |
|---|---|
| To: | Andres Freund <andres(at)anarazel(dot)de> |
| Cc: | Robert Haas <robertmhaas(at)gmail(dot)com>, Jeevan Ladhe <jeevan(dot)ladhe(at)enterprisedb(dot)com>, PostgreSQL Developers <pgsql-hackers(at)postgresql(dot)org> |
| Subject: | Re: Binary search in fmgr_isbuiltin() is a bottleneck. |
| Date: | 2017-09-28 18:59:45 |
| Message-ID: | 12940.1506625185@sss.pgh.pa.us |
| Views: | Raw Message | Whole Thread | Download mbox | Resend email |
| Thread: | |
| Lists: | pgsql-hackers |
I wrote:
> [ we should use an index array ]
Just to prove the point, I threw together the attached trivial test case,
which time-trials the existing fmgr_isbuiltin implementation against both
the proposed hash implementation and a simple index array. On my machine,
with a repeat count of 10000, I get
NOTICE: bsearch runtime 4234.087 ms
NOTICE: hash runtime 2542.636 ms
NOTICE: index runtime 165.184 ms
(These numbers are repeatable within 1% or so.)
It could be argued that trialling OIDs sequentially gives a bit of an
unfair advantage to the bsearch and index methods over the hash method,
because the former are going to suffer fewer cache misses that way.
But I don't see a randomized lookup order changing the conclusion much.
regards, tom lane
| Attachment | Content-Type | Size |
|---|---|---|
| testruntime.c | text/x-c | 3.8 KB |
| From | Date | Subject | |
|---|---|---|---|
| Next Message | Tom Lane | 2017-09-28 19:03:38 | Re: Domains and arrays and composites, oh my |
| Previous Message | Andrew Dunstan | 2017-09-28 18:58:00 | Re: Domains and arrays and composites, oh my |