Re: spreading disk I/O ,which is better??

Don Haney wrote:
>
> Wijbrand Pauw wrote:
> >
> > Hi,
> >
> > I have the following problem, I have a heavily used tablespace of 8Gb and IBM
> > SSA disks. I can't use logical volume striping.
> > I have a choise of making one tablespace on a 9Gb disk or use filesystem
> > striping on 4 disks of 2.2 Gb. The last option means making several datafiles
> > on the 4 disks. If the last option is the best? and if so, what is wise?
> > making a lot of datafiles (say 200Mb each) and file 1-4 on disk 1-4 , file 5-8
> > again on the 4 disk's and so on until I have 9Gb or just 4 datafiles on the 4
> > disk's .
> > Is the choice different when dealing with indexes or tables (indexes tend to
> > be accessed more randomly)?
> > The disk's are mirrored so reading will be fast but writing will be slower.
> > SSA disk's are supposed to be very fast, IBM tell's me a single SSA disk can
> > compete with striped FW SCSI disk's. I can make a group of disk's in which the
> > other disk's won't influence this disk, so the SSA adapter won't be a problem.
> > For people who know SAP , it's the sappcl2 data and index tablespace of the
> > payroll.
> >
> > Thanks for helping!
> >
> > regards,
> >
> > Wijbrand Pauw
>
> Wijbrand,
>
> It is always better to place your data and indices in different
> tablespaces and to place the data files that comprise the tablespaces on
> different disks, if possible. This will reduce disk I/O contention and
> should help performance. Since logical striping is not available, as in
> using RAID hardware, the striping performed by ORACLE is a little more
> tricky in that you have to set the file sizes so that the 1st, 2nd and
> consecutive extents are allocated on different disks. Unless you use
> MINEXTENTS 2 in your creation clause, you won't see any benefit at all
> until the 1st extent is filled. Each file would have to be slightly
> bigger than the initial extent to allow for growth.
>
> This method is a little tougher to maintain than logical striping. It
> is more enforced by the DBA than the Operating System or Oracle.
>
> Hope this helps,
>
> Don

Don,

I would use the the filesystem striping of the two choices you gave. I would set the degree of parallelism to #CPU's *2. and have the same # of datafiles in the tablespace.

I have done extensive I/O testing on an RS/6000 J30, 8 604e CPU's with SSA drives. The best performance came from using Raw partitions ( 54.6 MB/Sec, MCA Bus Bound ) compared to filesystem ( 32 MB/sec, CPU Bound ). Received on Tue Oct 14 1997 - 00:00:00 CDT