Re: Direct IO

;-)

The little I know about it, and in the context in which our earlier 'discussion' would have raised the issue, this has nothing to do with by-passing Oracle's buffer cache, but to do with by-passing the file system buffer.

NT and Raw devices don't use a file system buffer. Unix does (generally), and its buffer is usually 8K big (hence the need for Oracle to match that buffer, and thus need 8K blocks on Unix).

SORT_DIRECT_WRITES is a different beast entirely. That tells Oracle that when the PGA for a User has filled up doing a sort, and needs to swap to temporary tablespace, could you please not flood my buffer cache with the relevant data, but have the Server Process write it directly to the relevant datafile. Provided your sort_area_size is at least 640K in 8i, sort_direct_writes is automatically true (as far as I can tell).

But although the buffer cache is thereby avoided, you'd still actually be passing the write to the file system buffer if your temporary tablespace was on a cooked device.

Regards
HJR "Dave Haas" <davidh_at_no.spam.hotmail.com> wrote in message news:SERY6.338$84.100633_at_news0.telusplanet.net...
> Hi.
>
> I have a question with regard to terminology. In several posts (and an
> argument I had with Howard a while ago :) people have used the term
 'Direct
> IO'. To be perfectly honest I'm not exactly sure what that means. AFAIK
> the IO options are 1) file-system buffered and 2) Raw IO. I heard (or
 more
> to the point, read) a post that said something to the effect of '...
 direct
> I/O means that the buffer cache is not involved in the operation ...'.
 Does
> that have something to do with the sort-direct-write operation and the old
> SORT_WRITE_BUFFERS and SORT_WRITE_BUFFER_SIZE? I'm a little confused (as
> usual) ...
>
> Dave
>
>
Received on Mon Jun 25 2001 - 06:48:31 CDT