Author: David Dory
Date: 13:39:03 05/31/02
Go up one level in this thread
On May 31, 2002 at 11:16:32, Robert Henry Durrett wrote: >On May 31, 2002 at 01:19:57, David Dory wrote: > >>>You have some very interesting ideas in your post. But before getting to them, >>>I want to make sure we are "singing on the same sheet of music," so to speak. >>>My interest in this topic developed while thinking about the upcoming >DF/Kramnik match in October. I tried to imagine a configuration which would >have a reasonable chance of being built in time for that match. Since no one >has posted anything here about the computer that is to be used for that match, >I found myself speculating as to what could possibly be planned. The first >>>thought that came to me was: "Whatever it is, it's going to have to be >>>something they can build and deliver by August, if Kramnik is to have an >>>opportunity to play with it for a few weeks before the match." This led to >>>recall of the often quoted principle "Keep it Simple Stupid [KISS]." To me, >>>that meant that anything that would require extensive design, development, and >>>testing would be absolutely out of the question. This prompted the question >>>"What could be put together in just a few months?" Although non-optimal, the >>>configuration I asked about seemed to me to be one possible choice for that >>>match. >>============= >>This configuration is already available. It's called a "cluster". It has a >>number of independent m/board's fitted into a common bus (backplane). Each mobo >>works independently of the other, and has it's own memory, etc. > >That brings back memories. I have been monitoring this bulletin board off and >on for a number of years now. I vaguely recall bulletins discussing clusters. >Also saw clusters discussed in technical journals before I retired. [Don't look >at technical journals now since no motivation to do so.] > >At the risk of sounding like an irritating broken record, the key issue in my >mind is the speed of the microprocessors. [Also speed of the boards.] Although >the cluster configuration concept has been around for years, the clusters built >and evaluated always are limited by the availability of suitable components. A >cluster built two years ago necessarily had to use some comparatively slow >hardware, for example. It does no good to use a 100% efficient design if the >components have to be terribly slow! No. Clusters are FAST CPU's. There are no shared memory issues. Bandwidth between the processors is not a problem. You're referring to the shared memory multi-CPU computers. They do run slower, but taken together, with a good parallel program - they do romp! > >> >>The computer for the Kramnik match is (IIRC), already chosen. It's from Compaq. > >Forgive me, but I forgot what IIRC means. IIRC = if I remember correctly. (don't worry, I usually don't rc :-) ) >Are you sure? Some of these companies give the likes of Kasparov and Kramnik >millions of dollars just to get a little publicity! Crazy! I'm sure! They may let Chessbase have a go with their latest multi-cpu board, but they won't be bringing out some Itanium3. Maybe an Itanium2 (formerly named McKinley). >Don't be afraid to talk about information bandwidth and Shannon's theorems if >you want to. There are quite a few people here who can discuss them >intelligently. [I've seen the bulletins.] Back in the 70s, when I was working >on my MS degree in Electrical Engineering, my major was "Communication and >Information Theory." We studied all that good stuff then, but all the examples >and problems were in the context of communication systems. Bandwidth I know a tad about (and only a tad). Shannon's theorems I don't even have a clue about. >I guess we need an input here from someone more familiar with this. I don't >have a good feel for how long it would take. Perhaps finding the most attractive >eight or sixteen two-ply sequences would be sufficient. Move generation is done VERY fast. Not to worry. > That would save a >little time. Also, it may not be necessary to have a "high fidelity" >evaluation. A rough estimate might be sufficient. There are tradeoffs here. Yes, and certainly yes. You're batting 1,000! >>I'm not sure if PCI bus will support 400 MB/sec or not. Sounds high, maybe?? > >True. I picked 400 MHz as an "upper bound." > >Another "broken record": The only motivation for selecting a non-optimal >design >was to enable use of very fast processors. I suspected that the speed of the >processors was more important than the efficiency of the design. If fast >processors could be used, no one in his/her right mind would select anything >except the most efficient design. I went into a "deep think" on this (and this is always dangerous for a guy with a shallow head). And I think this idea (which as you can tell I've been playing with for a while, not willing to "GET WET", has more merit than I anticipated. A holy terror for more work though, in the coding. (at least for my coding) >Maybe that's in Rosch's new book, but I haven't seen it there yet. Just got the >book and haven't read it all. [1348 pages of text!!!!!!!] Little light reading, have we? <grin> >What I was considering was a "network" [or "bus"] which would be completely >enclosed by a standard 19" rack. The small size would facilitate speed. A much >larger network/bus might be a real mess, especially if messages had to be routed >through routers [and much more] as on the Internet. Probably totally >impractical if speed is the objective. Yes, physical proximity is a priority for the cluster. (hard to imagine that came out of my keyboard, actually) > >All "computer-chess geeks" surely are rooting for the computer. Right? > Both, actually. I just want to see a good match - no dumby "computer" moves, and a close score and I'm dancing on the ceiling! >There is a lot of information in Rosch's book about PCI. It will take me >awhile to digest it all. No problemo, amigo. Beano, anyone? Check this out: http://www.pcworld.com/news/article/0,aid,100544,00.asp What a story!! Dave
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.