Archives of the TeradataForum
Message Posted: Thu, 26 Sep 2002 @ 17:33:26 GMT
| Subj: | | Re: AMP Worker Tasks |
| |
| From: | | John Hall |
Thanks to Fred and Calvin for their information.
What led to my question was the thread on OLAP spool behavior that Jeremy Christiansen posted some time back. In that thread, Jeremy
observed that as the result of using some OLAP functions, he was seeing spool skew to just a few AMPs. I found his observation to be
interesting and thought about the point that even if I have otherwise well behaved queries, the use of those OLAP functions could lead to
transient hot-AMPs. A few short duration hot-AMPs probably aren't a big deal - but they are when I have hundreds of concurrent queries that
are experiencing varying degrees of hot-AMPing.
So if some of the OLAP functions are causing spool skew, then aren't they also likely to be causing long-lasting AWTs on those AMPs as
well? Once again, a few queries with that kind of behavior probably isn't a big thing, it's transient. A lot of queries behaving like that
in a mixed workload environment is an issue that starts to drive the need for an upgrade. And even after the upgrade is done, the full
value of the upgrade is probably not going to be realized. The best answer is having the maximum understanding of your workload and
minimizing those things that hurt performance.
So that led to my question. Currently, queries and workload are designed, tested and implemented. It's only when a process/query can be
identified as being a problem, after the fact, does it get attention. If I have some knowledge of the system demands involved with my
various workloads, then I'm in a better position to design new workload. If I ignore or am ignorant of the costs and inter-relationships,
then my odds of doing things right the first time are greatly reduced.
I think that having a better understanding of AWTs is important to improving the performance of the Teradata, hence my question. Knowing
the number of AWTs is certainly one side of the equation, but Fred is right that duration of an AWT is probably more important. It would
sure be nice if there was some documentation rather than having to guess at how the machine works.
| 
