Wildly inaccurate profiling results

I have a program, compiled with cost-center profiling on, that takes about 5
minutes, 30 seconds to run (measured with a stopwatch), but the resulting .prof
file reports the total time as 34 seconds. I know that this number doesn't
include overhead due to profiling, but surely the overhead can't be that much!
(Also, the program runs about as long when run without -P.)  I'm using a copy
of GHC built from sources checked out around July 1. I'd include the program,
but it's very large and I'm not sure how to minimize it and preserve the same
behavior. Any ideas as to what the problem might be?

Thanks,
Kirsten

-- 
Kirsten Chevalier * krc@xxxxxxxxxxxxxxxxxxx * Often in error, never in doubt