Performance Questions
 

Sorry to bombard the forum with basic questions, but I have found little on the internet regarding how Prime95 (or any of the GIMPS engines) make efficient use of the cache. I understand that without a cache the memory bandwidth requirements would be emornious. I know that an N-pt FFT can be divided into several smaller FFTs by arranging the data in a matrix form and performing FFTs on the rows and then (after a transpose) the columns. Does GIMPS use this?

Also, could someone point me to a performance analysis of the GIMPS algorithm? I'm interested in how much time is spent doing the various parts of the LL Test.

Thanks again!
Nevarcds

Did anyone ever get back to him?

[quote=nevarcds]how Prime95 (or any of the GIMPS engines) make efficient use of the cache.

< snip >

Does GIMPS use this?[/quote]
See George Woltman's notes at [URL="http://www.mersenne.org/gimps/p4notes.doc"]http://www.mersenne.org/gimps/p4notes.doc[/URL]

But be aware that they're five years old and don't include his later findings on P4 cache and SSE2 instructions.

[quote]I'm interested in how much time is spent doing the various parts of the LL Test.[/quote]
99% in the FFT routines, IIRC.

[QUOTE=cheesehead]99% in the FFT routines, IIRC.[/QUOTE]
No, it's more like 80-90% doing the FFTs, the remainder doing the dyadic-squaring and round-and-propagate-carries steps.