Re: Performance Question: RAID or Individual drives ?

MotoX wrote:
>
> Art S. Kagel wrote in message <35AA596B.464F_at_bloomberg.net>...

> >surpassed. This is exactly the behavior that one gets with modern
> >database parallelization such as Oracle and Informix provide.
>
> Not for a distribution of many, small random I/O's. Striping can actually
> make things worse. And in those processes are each requesting data from
> completely different areas of the RAID set - even sequentially - things can
> still get worse. That's why it's so important to test you own system with
> your own transactions, as I mentioned to the original poster.

You are right that a stripe set with large stripe size (64K) is bad for random reads. However, tests have shown that for smaller stripe sizes (say 16K) random performance is not as bad. (BTW 16K happens also to be the favorite I/O write request size for Informix, and I believe for Oracle also so this is a better match to the server's activity anyway.) In addition, keep in mind we are talking about a disk farm for a relational database not for a filesystem. Oracle and Informix have impressive buffer caches and do significant and intelligent read-ahead so that their disk accesses are rarely random at all! Even writes are coallesced so that writes are in large sequential blocks. Sorry but for Oracle and Informix RAID10 is just clearly superior, more expensive for sure but better. And for installations where cost is a factor RAID3 is superior to RAID5 for both performance and safety. Let the LVM do the RAID10 or RAID3 it is worth the administration overhead.  

> >It looks to me like Motox's testing is using a single Wide or Fast only
> >controller and is hitting his controller's I/O limit (10MB/s).
>
> Nope, SSA. 4 x 20MB/s ports, 2 loops, per adapter.

Sounds cool, but it does not explain your results, I see a bottleneck somewhere. Is the RAID set limited to a single controller or spread? Just curious.

Art S. Kagel Received on Tue Jul 14 1998 - 10:00:39 CDT