Author: Robert Hyatt
Date: 10:45:18 05/06/04
Go up one level in this thread
OK. First, my memory was faulty. What I ran on the opteron was the 24 CB positions from the DTS article. I had thought I ran the BK positions, and, in fact, I am sure I did, but the logs were not where I expected. I will continue to look. But these are probably better as Vincent thinks that Crafty does far worse on these than it did on the BK problems. So lets go with worst case for the moment. Next problem. I ran all 24 positions to a fixed time limit of 5 minutes. What you will have to do to get the actual times is look at the mt=0 (non-parallel) log file and for each position, look at the _last_ reported output (IE a PV change, an iteration completed, whatever happened _last_. Then look at the corresponding output in the mt=2 and mt=4 logs to get the times. mt=2 and mt=4 will search deeper so you have to be sure you compare apples to oranges. IE if mt=0 displays the PV for the first move at depth=14 and then runs out of time, you have to find the place where mt=2 and mt=4 produced that same first PV and grab the time, not caring if they actually searched and found deeper stuff. I did a quick sanity check and got 4848 seconds (adding up the times for all 24 positions) in the mt=0 log. I got 2569 seconds for the mt=2 log (overall speedup is therefore 1.89X) and I got 1634 for the mt=4 log, giving an overall speedup of 2.97. If your total times disagree, we need to reconcile. I could have made an error. The actual log files are available at ftp.cis.uab.edu/pub/hyatt/smpdata Of course you will want to compute the speedup for each position so that you can produce some sort of error estimate. I gave the above only as a quick look at what I saw... Let me know if there are questions. Note each of the 3 logs is about 100kb so they are not huge. I opened up mt=0 in one window and either mt=2 or 4 in a side-by-side window and searched for "correct" which is the last line (correct or incorrect) displayed for each of the 24 positions. That makes picking up the times pretty easy by direct comparison to be sure you look at the right line in the SMP logs... Bob Note that this was on the quad opteron 2.2ghz box that I actually used in the CCT event. It is identical to the released version with the same number except that the libNUMA stuff is not included for reasons given previously...
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.