Author: Robert Hyatt
Date: 08:39:04 01/15/06
Go up one level in this thread
On January 15, 2006 at 04:45:07, Gian-Carlo Pascutto wrote: >On January 14, 2006 at 18:28:39, Robert Hyatt wrote: > >>My parallel search is bug-free. > >I've saved this one for future amusement :) Sounds good to me. I define "bug" as "produces wrong answers" or "crashes". I don't see any of that, period. Not that it can't be improved, performance-wise. But performance isn't a "bug". > >>>Realistically, the SMP mode of most programs is a lot less tested than the >>>normal mode, because almost noone used to have a dual machine before the X2 >>>arrived. >> >>I don't buy that at all. I've been teaching parallel programming for many >>years. Students had dual-cpu machines 5 years ago. They were not particularly >>pricey, other than factoring the extra CPU. I've been running parallel on ICC >>since 1996 or so when the first dual-cpu machines started showing up, and my old >>quad pentium-pro 200 box was delivered. I think the commercial guys have quad >>boxes available for testing all the time... > >Huu, maybe in your ivory tower where hardware doesn't cost anything this is >true. But just look at this forum. Only 2 years ago basically *no-one* had a >dual machine. Today a *large* amount of people have X2's. That is 100% wrong. To prove the point, hopefully some that had duals prior to January 2004 will post here to illustrate. Students are not wealthy. I had a dual PII/300 in 1997 that sold for about what my pentium pro 200 home box sold for when Gateway first released those. Anthony (Zappa) for example has a dual dual-core (4 processors). I've gotten email from others with these boxes. But I've been getting SMP questions for 6-7 years now from people with duals, some with quads, a few with access to even bigger boxes. >These people run other tests, different configurations than the ones we run, and >they *will* find additional bugs if there are. > "if". It is impossible, as I mentioned, to prove that no bugs exist. It is possible to prove there is one, by just demonstrating it. But the former is impossible. As I mentioned previously, A program has a serial and a parallel version. It is just as likely that there are bugs in either version. Particularly since most of the code is unrelated to parallel search. It also happens that a parallel program produces correct answers while the serial version does not. So worrying about bugs in either is pointless when trying to answer the question "if a program runs 1.7X faster on a dual 1.0ghz box, how would that compare to a 1.7ghz box?" I could, and still would say "pretty equal" and move on to the next question. Otherwise you get way off into left field. Because program X might have a bug only triggered by deep depths possible only on fast hardware. And so the serial version on the faster hardware might be worse. This happened to me more than once in the Cray Blitz days, where the only problem was we could just go deeper than we thought possible, and blew out some fixed size data structures. If you go off on that tangent, how can you _ever_ say that a machine 2x faster is "better"??? But in general, deeper search is stronger than shallower search, all else being equal. So going 1.7X faster, whether it be via parallel search or faster single-processor machine doesn't really matter. Faster is better. >-- >GCP
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.