Archives of the TeradataForum

Message Posted: Fri, 11 Jun 2004 @ 22:02:53 GMT

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

The best reference for this subject is the Orange Book: "Understanding AMP Worker Tasks" I highly recommend that you read it. It is well written, and explains this subject in easy to understand language.

A couple of quotes from that book:

A fixed number of AWTs are pre-allocated at startup for each AMP in the con-figuration, with the default number being 80. All of the allocated AWTs can be active at the same time, sharing the CPUs and memory on the node.

Below is a list of some of those and how they use AWTs:

- ArcMain: One AWT per AMP.

- FastLoad: 2 AWTs per AMP during the acquisition phase, if number of sessions equal the number of AMPs, one as a sender AWT, one as a re-ceiver AWT. Otherwise the number of sessions will determine the number of AWTs in use across all AMPs for the sender work, while the number of receiver AWTs will equal 1 per AMP. The Apply phase always uses 1 AWT per AMP.

- MultiLoad: Same as FastLoad for the acquisition phase, 1 AWT per AMP per table being loaded on the apply phase.

- FastExport: However many AWTs the query requires will be used, plus the following after the query completes. The horizontal and the vertical partitioning steps each require 2 AWTs per AMP (1 sender and 1 re-ceiver). The export phase requires 1 AWT per AMP.

- TPump: TPump does not require AWTs to be held on each AMP during the duration of a job or a phase. Each individual TPump update ac-quires/releases its own AWT on a single AMP, holding the AWT for a very brief period of time. Each base row update may utilize more than 1 AWT if the row being updated has fallback, a unique secondary index, join in-dexes or triggers. The TPump pack factor times the number of sessions provides a ball park estimate of how many total AWTs may be required for the base update at any point in time.

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