You set your populations however you want. You hopefully true your internals first (as Hewitt hammered home), then sample until the picture emerges (or the money runs out).
Morris did a monster DvR national LV sample in May that I have a lot of respect for:
"From May 5-11, 2012, I conducted a survey of 6,000 likely voters. On such a mammoth sample, the margin of error is less than 1 percent. I found that Romney has amassed a sizable lead over Obama of 51-42, far in excess of what published polling and surveys of registered as opposed to likely voters are indicating."
This post was so prescient...