gin uses maintenance_work_mem,so try to increase it and see dramatic

On Mon, 26 Jun 2006, Stefan Kaltenbrunner wrote:

on IRC somebody mentioned that it took >34h to greate a GIN index (on a
tsvector) on a ~3 Million column table (wikipedia dump) with a
reasonable speced box (AMD 3400+).
After getting hold of a dump of said table (around 4,1GB in size) I
managed to get the following timings:

test=# CREATE INDEX idxFTI_idx ON wikipedia USING gist(vector);
Time: 416122.896 ms

so about 7 minutes - sounds very reasonable

test=# CREATE INDEX idxFTI2_idx ON wikipedia USING gin(vector);
Time: 52681605.101 ms

ouch - that makes for a whoppy 14,6hours(!). During that time the box is
completely CPU bottlenecked and during virtually no IO at all - (varing
maintainance_work_mem does not seem to make any noticable difference).

That box is a fast Dual Opteron 2.6Ghz with 8GB RAM and a 4 disk RAID10
for the WAL and 12 disks for the data running a very recent -HEAD
checkout ...

It looks like we still don't have any docs for GIN in the tree so I
don't know if those timings are expected or not ...


---------------------------(end of broadcast)---------------------------
TIP 2: Don't 'kill -9' the postmaster

Oleg Bartunov, Research Scientist, Head of AstroNet (,
Sternberg Astronomical Institute, Moscow University, Russia
phone: +007(495)939-16-83, +007(495)939-23-83

---------------------------(end of broadcast)---------------------------
TIP 2: Don't 'kill -9' the postmaster

Reply via email to