Ed Jaffe wrote <begin extract> Most of that is non-sequitur. We've already established in prior conversations that uncached memory access is approaching 1000 cycles on modern System z machines. That's roughly 75 times slower (relatively speaking) than memory access speed on the Model 91. </end extract>
I agree with EJ that cycles are a convenient and entirely comparison unit here These discussions are, however, problematic in another way, The cache and other such programmer-inaccessible machinery are devices for optimizing and in particular speeding up the code that programmers write or translators generate. Their characteristics, mostly but not entirely undocumented, must be inferred, at least by programmers outside IBM, from black-box experiments, In my own experience these experiments yield differing results over time: IBM is changing/improving this machinery at intervals that are short enough to be, in part at least, detectable. Those of us who write assembly language are shooting at a moving target. An optimal exploitation of the characteristics of this 'secret' machinery today may well be suboptimal tomorrow. Moreover again, it is already clear that some horrendous maximal cycle-time values, while entirely correct, are larger than life, even misleading. Statistical measures of central tendency and variability about them are needed. All of this has made me suspicious of very specialized attempts to exploit the current characteristics of the cache. Something can of course be done, and certain rules of thumb are likely to remain useful. Locality of reference was always a good notion, and now it is a crucially important one. John Gilmore, Ashland, MA 01721 - USA ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
