Are you using the table datatype to essentially create tables in memory?

I ran into huge performance problems with these because they cannot be indexed and are always accessed via a table scan. They are great for building result sets which are then downloaded all at once, but no good for OLTP stuff.

This may be totally irrelevant to you.

The other thing that occurred to me is: are you doing any sort of interaction with your client during this processing? If you're client is seeing a progress meter, your problem could be one of network traffic?

When you say temporary tables, are these actually permanent tables that you truncate when the process starts? Doing it that way means that you can index appropriately and if necessary place the table on a different drive set.

Talking of which, do you use delete or truncate to remove records? Are there triggers involved on the tables receiving the data results?

Regards
Simon