[Larceny-users] profiling spirales

[William D Clinger]

Eduardo Cavazos wrote:
> So what do the numbers mean? There are two tables
> which are output each time.

Roughly speaking, the first table shows the percentage
of time each procedure is executing its own code.  The
second table shows the percentage of time each procedure
is executing its own code or is suspended while waiting
for the result of some procedure it calls.  Hence the
percentages for the second table will always be at least
as high as for the first table.

Your profiles show that about 20% of the time is being
spent in %flonum->bignum.  I doubt whether your problem
really needs any bignum arithmetic at all, so I suspect
that you are using exact numbers for calculations that
were intended to use inexact numbers.

That, I think, is the most likely reason Larceny is
running slower than Ypsilon.  Interpreted systems
often perform well on benchmarks that do a lot of
bignum arithmetic, because there is no interpretation
involved during the bignum calculations themselves.
Furthermore Larceny's bignum arithmetic	is known to be
especially slow, so an interpreter can easily be faster
than Larceny on	a bignum-intensive benchmark.

I'd suggest you go through your code and convert every
arithmetic operation that should be taking inexact
arguments into the corresponding flonum-specific
operation.  For example, change + to fl+.  Then you'll
get an exception every time an exact argument is passed
to one of the flonum-specific operations, and that will
make it easier for you to track down the exact numbers
that (in all likelihood) are making the program run so
slowly.

The flonum-specific operations should be significantly
faster than the generic operations, so the transformed
program (after you have debugged it) should run faster
in both Larceny and in Ypsilon.

Will