Redirected to 'the proper list' - pgsql-hackers@postgresql.org
On Wed, 11 Mar 1998, Chris Albertson wrote:
> Also, is anyone working on storage mangers? I was thinking that
> a raw partition manager would be good to have. Could be faster
> then one that uses the file system. Give it two partitions and
> it could do stripping and gain some real speed.
stripping can be done from the operating system level to give you
that 'boost'...and Oracle, in fact, moved away from the raw partition
level to just using the Unix file system...I believe it would
overcomplicate the backend, and give a negligible boost in performance, if
we had to build a 'low level drive interface'...
Actually Oracle *is* suggesting the use of RAW devices for heavy IO databases.
RAW devices have the advantage that the operating system does not cache
the pages. This can at once be a drastic disadvantage if the database does not
have it's own read/write optimization (only read/write whole pages or better multiples thereof)
an intelligent buffer cache and intelligent read ahead (here only the db engine can be really intelligent about it).
If you use file system files and have the above features in the database engine,
then all pages will be in memory twice, once in the file system buffer cache and
once in the database cache. This results in only half of the pages beeing cached and also
adds the additional overhead of 2 memcpy's.
I think the first step in the direction of RAW devices needs to be the implementation of the
tablespace idea, i.e. more than one table or index in one file. On the driver side,
if there is a logical volume manager, then accessing a raw device is the same as
accessing a file system file (read/write/seek).
Andreas