Re: [HACKERS] Problem with the numbers I reported yesterday (fwd)

Bruce Momjian wrote:
> >
> > I ran my performance tests some more times and it seems the numbers are not
> > really comparable. When I run PostgreSQL without -F I get a sync after every
> > insert. With -F I get no sync at all as all inserts fit well into the
> > buffer. However, Oracle in comparison does sync. Simply hearing the disk
> > access it seems as if they sync every two or three seconds.
> >
> > Does anyone know a way to really check both DBMSs?
>
> Many dbms's do buffered logging, that is they sync after the buffer gets
> full or after a minute or so.  We have the logic to add buffered logging
> to PostgreSQL and will be doing it later.  Right now, we only have
> non-buffered logging, and no logging.

Will there be (or is there) a provision for using raw disk devices,
rather than going through the UNIX filesystem?  This might be able to
reduce some of the overhead.  Also, if a transaction log were appended
serially to a seperate raw disk, there would be verry little seeking
needed and this (the log) could easily be fsync'd for each transaction
while the data itself is buffered.

Of course not everyone has two raw disks to devote to a database (or
even one raw disk), but for those who do, this could provide speed and
reliability.

Ocie Mitchell