Thanks to everyone who contributed. What a great resource the J Forum is! http://www.jsoftware.com/jwiki/Essays/Sort_Times summarizes the findings and provides more details on the code. There are 3 distinct groups in the ratios: over 2, around 1.7, and around 1.
I consulted my local friendly hardware guru AKA RBE AKA Bob Bernecky, and he said a bunch of things, the most relevant of which seems to be: Now, there is another little-advertised aspect of cache memories that may be doing us in here. Repeated, back-to-back stores into the same cache word (or line - I'm not sure exactly...) can cause stalls. Makes sense. You make the little word work hard (in C, z[*x++]++ where *x is the same value) and it gets tired. :-) ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
