> NEC-2 uses four sequential access files rather than the single direct
> access file used by NEC-4, so four times as much disk space will be
> needed with NEC-2. System virtual memory will probably be a better
> choice in this case if it is available.
>
> Jerry Burke
> LLNL
I must disagree. Block matrix factorization (as does NEC-2 with its 4 files)
is much faster than complete matrix factorization using system virtual memory.
Using a simple test (only 500 elements) with NEC-2, I found execution times:
270s running in RAM (4M to store the matrix)
>1hr running in virtual memory (giving the process only 2M)
1200s using the four swap-files (also in 2M, smaller MAXMAT)
I did not wait longer than 1 hour for the second test to end.
This behavior is understandable, since when operating on the complete matrix,
all of memory has to be traversed many times (no matter how you choose the
loops, or whether you transpose the matrix), leading to never-ending disk
paging. Working on the blocks, however, can be done in core.
Greetings,
Jos Bergervoet,
not the USA (if someone sends me NEC-4 I'll gladly test that too :-)
Received on Thu Oct 24 1996 - 08:31:38 EDT
This archive was generated by hypermail 2.2.0 : Sat Oct 02 2010 - 00:10:37 EDT