Archives of the TeradataForum
Message Posted: Wed, 25 Sep 2002 @ 22:03:45 GMT
| Subj: | | Re: AMP Worker Tasks |
| |
| From: | | Fred W Pluebell |
In simplest terms, a step is performed by one or more AWTs. As soon as an AWT finishes one step, it's available to start working
on another step, which may be part of a different request from a different session. So an "AWT bottleneck" is a symptom that
typically is caused by too many concurrent long-running steps. You might start by monitoring lock conflicts - blocked steps tie up
AWTs. Also look at spool usage to see if there's evidence of large redistributions. The "view SQL for an active session" feature in
V2R4.1 and later may also help you identify very long-running steps. And you want to minimize rollbacks.
The number of steps in an EXPLAIN or even the number of AWTs per step isn't nearly as important to this process as understanding
how long each step ties up the AWT(s). But to address the original question, here are some guidelines:
A single-AMP step requires the services of one or more AWTs on a single AMP; an all-AMP step requires AWTs on every AMP, etc. If
multiple steps are to be executed in parallel, each needs its own AWT(s).
Simple "local" processing uses a single AWT per involved AMP. Redistribution or Duplication will require a "sender" AWT plus a
"receiver" AWT (on each AMP). Table update may require an AWT on the Primary AMP plus one on the Fallback AMP, etc.
| 
