Hey Jeff -
It's hard to say what's going on here without looking at the logs. You mentioned Dataserver and "Local" data sources. By local, do you mean "embedded in the workbook you publish", or do you mean you're testing viz exection from Desktop using local (on the Windows machine) file / extracts?
I'd embedded extracts to be faster than dataserver extracts, and "completely" local to be faster than either.
I'm talking about embedded extracts. What's interesting is that the tde for the embedded extracts and published datasources I think are both stored in the same place (in the dataengine folder), but the published ones have to go through the sqlproxy dataserver and I see a lot more latency with this approach and I'm not sure why. Does the dataserver grab the entire extract and place it into temp somewhere for ingestion by the dataengine or just what is needed for the viz? And I'm not sure what causes the additional latency, but based on looking at the logs it appears to based on the amount of columns in the datasource, # of rows, and # of connected published datasources for the first dashboard to render.
Do you know of any performance benchmark comparisons forr this? We use published extracts a lot as we can extract once for a given subject area, and the extract usually has a few hundred columns to serve the subject area. Some of our popular ones have 50-100 workbooks hooked up. It's kind of like a double edged sword as it seems to save on repetitively extracting the data, but then has a performance hit on end user rendering as far as I can see.