Archives of the TeradataForum

Message Posted: Sat, 28 Jun 2003 @ 16:49:23 GMT

	<Prev	Next>		<<First	<Prev	Next>	Last>>

Charlie,

Something's fishy here. The story's passed through a few hands, so I don't quite know where it went wrong...

First of all, all tables occupy all AMPs. Our architecture fully exploits this "striping" and that's how we achieve the parallelism. Now what may happen is that the data gets "skewed" either by a "poor" choice of its primary index, so that all of the rows wind up on one AMP (the table still occupies all AMPs, but most of the space occupied by the table is on the one AMP). When this happens, a full table scan on that table will run primarily on that AMP, i.e., will not be parallel. In this case, the full table scan will take much longer.

The remedy in this case is to choose a primary index which distributes the data more evenly. The Teradata functions HASHAMP, HASHBUCKET, and HASHROW can help you evaluate primary index options. The hashing algorithm is pretty good and it's quite unusual that a nearly unique primary index will be sufficiently "skewed" to produce the above effects.

Hope this helps.

Pat Belle
Certified Master

	<Prev	Next>		<<First	<Prev	Next>	Last>>