Thread: Improved Load Speed From Current Sources

Dear List,

I have been having a bit of a play with using COPY to load a reasonably large
table - 3 million rows ~ 57000 8K pages ("fact0" for those of you that have
suffered these postings previously).

I decided to compare 7.1.2 and current (7.2.dev 29/0701)sources on a number of
different platforms.

The results show a definite improvement for the current sources on all tested
platforms :


Linux 2.4( Intel 266Mhz )    7.1.2   9m30s
..........................  7.2.dev 7m30s
Solaris 2.8( Sparc 2x450Mhz) 7.1.2   9m
..........................  7.2.dev 7m
Tru64 5.0( Alpha 2x500Mhz)   7.1.2   6m
.........................   7.2.dev 4m15s

The platform results are quite interesting themselves - surprisingly good
performance on the little Pc...

The Alpha  is a 2 year old 910...*not too sure about Alpha model numbers*...
The Sparc is a new E220

Postgresql was compiled with the native compiler on each platform. All
databases had fsync=true,shared_buffers=4000,wal_files=10 and wal_buffers=20

regards

Mark

<markir@slingshot.co.nz> writes:
> I have been having a bit of a play with using COPY to load a reasonably large
> table - 3 million rows ~ 57000 8K pages ("fact0" for those of you that have
> suffered these postings previously).

> I decided to compare 7.1.2 and current (7.2.dev 29/0701)sources on a number of
> different platforms.

> The results show a definite improvement for the current sources on all tested
> platforms :

Cool...

I suspect this must be due to the changes I made to avoid unnecessary
locking and lseek() overhead during insertion of successive tuples into
the same disk page.  I wouldn't have guessed it'd make that much
difference for a single backend, however.

            regards, tom lane