Author: Robert Hyatt
Date: 17:06:36 07/15/03
Go up one level in this thread
On July 15, 2003 at 17:14:45, Gerd Isenberg wrote: >On July 15, 2003 at 09:33:39, Robert Hyatt wrote: > >>On July 15, 2003 at 06:24:58, Vincent Diepeveen wrote: >> >>>On July 14, 2003 at 16:07:27, Robert Hyatt wrote: >>> >>>You measure the latency with those benches of sequential reads. >> >>No. lm-bench does _random_ reads and computes the _random-access_ >>latency. >> >>Don't know why you have a problem grasping that. >> >> >>>So already opened cache lines you can get data faster from than >>>random reads to memory. >> >>That also makes no sense. Perhaps you mean "already opened memory >>rows"? >> >> >>> >>>Random reads to memory are about 280 ns at single cpu P4 and about 400ns at dual >>>P4s. >> >>No they aren't. >> > >Bob, i found nothing wrong with Vincent's code. He does N-random hashreads and >aggregates the time used. I thought about some factor 2 error - but found no one >so far. Random Hashreads, like chess programs do. > >1e9 random hash reads take 265 seconds (including ~60 seconds overhead) on my >athlon-pc, however latency is defined. Any explanation? Any systematical error >or assumption? What does lm-bench do, to measure latency? > >Regards, >Gerd It is possible to cause _other_ problems. IE you can push the instructions in the loop out of cache, for one thing. There are others. The best numbers I have seen come from lm-bench. It was not a quick and dirty program, it has a lot of research behind it to address specific issues that were pointed out over a period of a year. It is very easy to use a "low impedence probe" if you know what that means. It actually affects the circuit it is measuring. 200+ns seems way high to me, when the chip latency is less than 1/3 of that. again, I'd run lm-bench on your box to see what it says, then you have to reconcile the differences.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.