 |
 |
Archives of the TeradataForum
Message Posted: Fri, 11 Jun 2004 @ 22:56:10 GMT
| Subj: | | Re: AMP worker tasks |
| |
| From: | | Ballinger, Carrie |
Just to add a couple of pieces of information to Tom Greene's earlier comments on this thread...
Although 80 AMP worker tasks are allocated on each AMP at start up, some number are reserved for special types of processing (abort processing,
for example). Practically speaking your typical AMP has about 62 AWTs available to service user work at any point in time. This finite number is
an intentional part of the internal work flow management inside Teradata.
Because AWTs are only engaged when a step is dispatched from the PE to the AMP, and freed up at the end of each query step, the average multi-
step query may engage several different AWTs during its execution, some for very short amounts of time, some for longer times, releasing them as
each step completes. But because an AWT is only assigned when an AMP step is sent from the dispatcher, an idle session, one with no current
requests active on the AMPs, would not be holding an AWT.
I've seen some success using TDQM workload limit rules in V2R5 to throttle back lower priority work, when the volume of such work is
contributing to AWT exhaustion, as a method of reducing demand for AWTs. Whether this works for you will depend on whether the category of work
putting the pressure on the AWTs can tolerate some delays, and whether you are on V2R5 and have access to TDQM. There is some information in V2R5
priority scheduler monitor output that shows AWT queue lengths by allocation group, called "Avg queue length". If you monitor that information,
it would give you a better sense of where the highest demand for AWTs is coming from. Then you could define a TDQM workload limit based on the
associated with that allocation group, so that queries beyond a specified number that originate from that performance group would be placed in a
delay queue. It may be that you'd see queries delayed only at certain times of day, and the delay may be for a short time. There's a bit of
trial and error involved in finding the right workload limit value, but the technique is fairly simple and effective.
Thanks, --Carrie
| |
