Connor McDonald <connor_mcdonald_at_yahoo.com> wrote in
news:3E93CCEE.6ABF_at_yahoo.com:
er getting sole access to the devices involved.
>
> Nowadays you've got sort-merge, index full scan, fast full scan, 100's
> of users, storage abstracted out to n-layers, file system read-ahead,
> the list goes on. The odds of getting the benefit of reference
> locality (ie the next physical op picking up right where the previous
> one left off) is virtually nil, possibly with the exception of redo
> logs. Thus the separation of data and indexes is probably a moot
> issue
I would disagree with this. There's no cost involved in separating indexes
from their tables and there still is the possibility that doing so can
reduce disk head contention. Sure, we've got many more join operations now
than just nested loops, but the optimizer still picks that at times and you
still will have times where the index blocks and table blocks that you're
after are on the same spindle regardless of how much striping is involved.
RAID, storage abstraction, and new join operations have certainly reduced
the possibility of disk head contention, but they have not completely
eliminated it.
Received on Fri Apr 11 2003 - 08:36:53 CDT
