"Merlin Moncure" <merlin.moncure@rcsonline.com> writes:
> ok, I've been in crunching profile profile graphs, and so far have been
> only been able to draw following conclusions.
> For bulk, 'in-transaction' insert:
> 1. win32 is slower than linux. win32 time for each insert grows with #
> inserts in xact, linux does not (or grows much slower). Win32 starts
> out about 3x slower and grows to 10x slower after 250k inserts.
Just to be clear: what you were testing was
BEGIN;
INSERT ... VALUES (...);
repeat insert many times
COMMIT;
with each statement issued as a separate PQexec() operation, correct?
Was this set up as a psql script, or specialized C code? (If a psql
script, I wonder whether it's psql that's chewing the time.)
> 2. ran a 50k profile vs. 250k profile. Nothing jumps out as being
> slower or faster: most time is spent in yyparse on either side. From
> this my preliminary conclusion is that there is something going on in
> the win32 api which is not showing in the profile.
Hmm. Client/server data transport maybe? It would be interesting to
try inserting the same data in other ways:
* COPY from client
* COPY from disk file
* INSERT/SELECT from another table
and see whether you see a similar slowdown.
regards, tom lane