Re: NEC-LIST: NEC Execution Times Under Windows

On Tue, 7 Oct 1997, Kintronic Labs wrote:

> We are running NEC4.1 on a PC and are surprised about the execution
> time results. Using the same executable compiled under Lahey F77
> -EM/32, we are seeing much less of an improvement than we expected.
> The execution times are only a factor of ~3 faster on a 200 MHz
> Pentium MMX With 128 MB of SDRAM (10 ns) in DOS box on Win 95 OSR2
> than on a 90 MHz Pentium with 16 MB of RAM (70 ns) in DOS box on Win
> 3.1. We expected an improvement closer to 10 times faster.

I would expect a >2x improvement because of the CPU and SDRAM. I
don't use Lahey, but I would expect a 32-bit number-crunching program
to do more or less the same under a DOS extender, win31, and win95.
The big factor to consider is the memory.

> The 3 times improvement factor is consistent on smaller ("in core")
> runs or larger runs which require swapping to disk (using virtual
> memory).
>
> It seems like the executables are not seeing the extra memory. Even
> the large jobs should not need to swap to disk.

Are you adjusting MAXMAT in NECPAR.INC accordingly? When you say
virtual memory, I assume you mean the out-of-core solution, which
works a lot better than virtual memory in the other sense (swapfile).

> Any suggestions if we were expecting too much. Or if we need a
> compiler upgrade. Also, we're going to try it under Linux as well.
> Any compiler suggestions would be appreciated.

You can compile it with g77 (GNU's "free" compiler) with -o3, but it
will probably be slower than what you're getting now, on a similar
system. I compiled it successfully with a beta of NAG FORTRAN 95
(using lots of compiler switches to suppress errors and warnings), but
the beta timed out before I did any benchmarks, and I can't find
anybody to sell me a copy.

I have a mystery of my own. I ran a 1924 segment model on nec 4, on a
large system with no swapping, with different values of MAXMAT. I
haven't finished checking the results, but it actually took longer
with MAXMAT=2000 than with MAXMAT=400. This is nothing like what
happened when I ran the same file using NEC 3 and varying IRESRV. I
have to play with it some more before I can ask intelligent questions,
I'm only mentioning it now to show that I don't have it all figured
out either.

Back to linux; if you run it on a linux system with more than 64 meg RAM
and get lots of swapping, make sure that linux knows you have more
than 64 meg. I had to add

                     append="mem=96M"

to my lilo.conf
Received on Fri Oct 10 1997 - 08:59:37 EDT