Author: Vincent Diepeveen
Date: 12:46:50 06/29/04
Go up one level in this thread
On June 29, 2004 at 12:42:41, Joshua Shriver wrote: >Actually I have an engine somewhat ready now.. reason I didnt enter it this year >is because I *AM* spending a lot of time (and most of next year) writing highly >parallized code. > >Yes there are supercomputers in Sweden. A friend of mine is the CEO of >Building31, who has a nice opteron cluster. (not sure of the # of nodes >offhand). > > >You main point was parallel coding.. I agree.. >If 10 ppl where given Quad Xeons or Quad Opterons, and their code wasn't >parallized vs ppl who wrote parallel code. Yes I would expect the parallel >*CODE* to win.. and would hope so since they spent the time to make it like >that. > >But that just brings me to my point. It's the CODE that wins.. not necessarily >the hardware. >-Joshua Shriver And at not a single tournament you get state of the art hardware for free. In graz we could get 2.4Ghz P4's. Of course nice the organisation provided them, but they were not sold in Netherlands anymore by then as they were too slow. Now you can get 2.8Ghz prescott cpu's for free in 2004 world champs. Very nice from organisation to give me one (needed to remote connect of course), however they are not getting sold here anymore, that slow they are. In short you do not parallellize your code to not show up with a parallel machine. Even if it was 40% slower than the machine i connect remote to, i would take the local machine, especially for the blitz. However you can see at aceshardware.com that single threaded a single cpu A64 is already 40-50% faster than a 3.4Ghz P4 EE. That's just 32 bits. My very outdated machine at home is a dual K7 2.127Ghz. Effectively it is a lot faster than a single cpu P4 2.8Ghz prescott. Let's say factor 2.2 or so? If you want equal hardware challenges, then also realize the consequence, that's that you will be playing at outdated hardware. Do you guess that the average audience is more interested in outdated hardware, or in hearing how state of the art machines do, especially when playing your big opteron cluster. By the way, the latency at an opteron cluster using cray technology (XD1) is a 1.5 usec one way pingpong latency One way ping pong latency at network cards put in dual opteron nodes (either selfmade cluster or from IBM or something) is about 3.5 usec latency. One way ping pong latency at the TERAS supercomputer i played at with 500 processors is about 3-4 usec one way pingpong latency. It was 5.8 usec on average to get 8 bytes at 460 processors (all 460 processors simultaneously requesting 8 bytes from other processors at random) using in total 115 GB ram for the random requests. So parallellizing at the opteron cluster you will face similar problems i had at the TERAS supercomputer. Hopefully you can test at it. I know already now know that it is unlikely you get something useful to work before 2005 WCCC at this machine as you will be busy night and day from now on getting to work something at it. Yet everyone will be very HAPPY if you join at it, because you will raise the esteem of the tournament for the average user. Your 2Ghz P4 tournament won't do that.
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.