Readers solve Pentium III laws of physics dilemma
Cache for questions
You may not want an Intel Coppermine Pentium III. You might well be better off with an older Katmai processor instead. For although the newer part boasts a level two cache running at full core speed, it is now clear that despite Chipzilla's claims that 256K running at full speed is equal to 512K at half speed, Coppermine is outperformed by the older chip when handling large amounts of data, and outperformed by slower chips to boot. When we revealed that a 700MHz Coppermine performed worse than a 500MHz Katmai when running SETI at home analysis (Story: Pentium III defies the laws of physics), we were at a loss initially to explain how this could be. But Register readers came to the rescue and supplied an answer that in retrospect is so obvious that we should have spotted it ourselves. Thanks to everyone who wrote in. Kenneth J. Hendrickson of Florida takes up the story: "There will be some applications where the data fits entirely into the smaller cache on the 700 MHz Coppermine, and for those applications it will appear faster. There will be other applications where the data is too large to fit into either processor's cache; in that case, factors other than CPU speed will determine performance. "The fastest processor in the world is not impressive unless you can keep it fed with data. Perhaps the Coppermine is starved for data in the SETI application." Guy Morrogh adds: "One processor runs all the system processes and handles all the interrupts, including handlers for the clock, disk, screen (drawing graphics etc). These have to be swapped into the cache, meaning that the benchmark gets pushed out and has to be reloaded into the cache. With two processors, only one processor gets hit when these run. "SETI likes a bigger cache - you'll probably find it works better with a larger cache, even if it is slower. If it overflows a 256K cache, but not a 512K cache, it will have to make a lot of hits to main memory, which is going to take a lot of clock ticks with a 7x clock ratio." "I can perhaps help you with the Lancewood mystery," Dean Johnson says. "I have a SGI 1200 with dual 700e's and 1Gb of memory and most of the time it is running SETI@Home, for which I have the source because I am on the porting team. "Under Linux, the SETI@Home client uses about 13Mb of memory, so the overall memory isn't an important thing. The key, however, is the size of the cache. The data for a Seti work unit is about 350K, so the difference between smaller and larger caces is huge due to the poorly optimised FFT (fast fourier transform) that eats up most of the CPU time. For instance, A PII 350 is twice as fast as a Celeron 466. So if you're manipulating large wodges of data, you're better off with an old 500MHz Katmai PIII than a shiny new 700+MHz Coppermine. And in dual processor configurations, the older chips benefit from having a whopping 1Mb of L2 cache, meaning they hardly ever suffer a cache miss even with the most data-hungry applications ®
Sponsored: RAID: End of an era?