 |
 |
Archives of the TeradataForum
Message Posted: Fri, 01 Nov 2003 @ 01:45:43 GMT
| Subj: | | Re: Number of sessions for mload / fastload |
| |
| From: | | rmhsmd |
It depends. Traffic flows on these sessions during the acquisition phase of FDL/MDL. The data sessions communicate directly with
the AMPs, NOT via the PEs. The TCP/IP connections to the Teradata Message Subsystem are evenly balanced across known gateway IP
addresses; Teradata somehow assigns each session of a job to a different AMP. The amp receives the data, determines the row hash,
and does buffered row redistribution to the owning amp during the acquisition phase. Thus all amps are doing something and the amps
with sessions assigned are doing more.
I focus on 2 main considerations in choosing the amount of sessions: Throughput and balance. A graph of throughput v session
count would show throughput initially increasing with session count, then leveling off to where additional sessions don't improve
throughput because there is some other bottleneck. By balance I mean that too few sessions can skew the work done by the nodes
containing the amps assigned to those sessions.
| |
