Archives of the TeradataForum

Message Posted: Tue, 18 Nov 2003 @ 10:06:28 GMT

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

Hi Padhu,

Fastload logs on multiple sessions - typically one per AMP. When sending data to Teradata in Phase 1, Fastload sends multiple rows at a time, referred to below as a 'transfer block'. I think the 'transfer block' is 32KB.

It reads the input file and packs as many data rows as possible into each 'transfer block'. When the first 'transfer block' is full it gets sent to the DBMS using session #1. Whilst the DBMS is processing this first 'transfer block', the Fastload utility is building the second 'transfer block', when ready that gets sent to the DBMS on session#2. This process continues, sending each 'transfer block' down separate sessions on a round robin basis, until all the input data has been read. This is Phase 1 of Fastload.

Then Phase 2 starts which results in the rows being sorted into row hash order and making the table available for queries.

Is this double I/O and an overhead ? Not sure, the alternative that I see it is to separate the input records into specific AMP 'transfer blocks' on the client and then send them to the correct AMP. This would mean potentially large emory and processing requirements on the client side. BUT even if this was done, the rows would still ned to be sorted at the DBMS end because there is no gaurantee that the records would be sent to Teradata in row hash order. Just my opinion.

Cheers,

Dave

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