Author: Robert Hyatt
Date: 08:09:32 09/06/02
Go up one level in this thread
I just created a new sub-directory on the ftp machine.. ftp.cis.uab.edu/pub/hyatt/smpdata There is a read.me that explains what the files are. I have uploaded the 1cpu log, the four 2cpu logs, the raw data file I created, and the program that will take that raw data and compute the speedups and print them in a readable fashion. If I made any errors in going from the logs to my raw data file, please post the "correction" here and I'll update the table and post it again... I found a couple of errors last night between posting the three 2cpu table and posting the four 2cpu table. I corrected the errors. But that doesn't mean there are not others. I am going to re-check myself as well, but you can draw any conclusions you want from the logs. They might convince you that SMP can turn your hair grey at times. :) For more fun look at the data I posted for Martin _first_. One position ran _slower_ with two cpus than with one, once, and got zero speedup in another case, and also had two that were just over 2.0... wildly varying... The couple of 2cpu positions that produce > 2.0 speedups every run need some attention. That indicates that for some reason, my move ordering is simply going badly wrong. But since it is repeatable, it ought to be findable as well. Whether it is actually "fixable" is unknown, because a deeper search _should_ reveal things about a position you didn't know after completing a shallower search, and all the move ordering "tricks" might be loaded with biased data that is suddenly horribly wrong... Also the positions that consistently produce speedups of less than 1.25 or so also need equal attention as those are even more problematic, since a lot of work is being done without much return. The position with a negative speedup (slowdown) is going to get a lot of attention from me (the position kopec 2 that I first posted for Martin last night). :)
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.