Author: Robert Hyatt
Date: 19:55:44 08/21/02
Go up one level in this thread
On August 21, 2002 at 18:32:21, Vincent Diepeveen wrote: >On August 21, 2002 at 17:55:44, Robert Hyatt wrote: > >>On August 21, 2002 at 17:32:11, Gian-Carlo Pascutto wrote: >> >>>On August 21, 2002 at 17:25:23, Robert Hyatt wrote: >>> >>>>And it is just as accurate >>>>as trying to count nodes. >>> >>>It isn't, because the NPS changes, which is just what I said. >>> >>>-- >>>GCP >> >> >>It doesn't change _that_ much on the last 3-4-5-6 searches. At least it >>doesn't for me in the results from Cray Blitz on 16 and 32 cpus... Once > >Let's jump onto this nodes a second. > >Let me quote some text written by you. > >DTS article 1997, march 1997. Issue 1 advances in computerchess. >Not exactly a CCC a bit way more official. What is this? I have no idea. My DTS article was only published in the JICCA. > >page 16 table 3 and 4: > 16 processors 155,514,410 nodes needed: 311 seconds = 500046 nodes > a second > 8 processors 109,467,495 nodes needed: 274 seconds = 399516.4 nodes > a second. > I can't find that data in my DTS article. Here is the node counts from table 4: pos 1 2 4 8 16 1 2,830 1,415 832 435 311 2 2,849 1,424 791 438 274 3 3,274 1,637 884 467 239 4 2,308 1,154 591 349 208 5 1,584 792 440 243 178 6 4,294 2,147 1,160 670 452 7 1,888 993 524 273 187 8 7,275 3,637 1,966 1,039 680 9 3,940 1,970 1,094 635 398 10 2,431 1,215 639 333 187 11 3,062 1,531 827 425 247 12 2,518 1,325 662 364 219 13 2,131 1,121 560 313 192 14 1,871 935 534 296 191 15 2,648 1,324 715 378 243 16 2,347 1,235 601 321 182 17 4,884 2,872 1,878 1,085 814 18 646 358 222 124 84 19 2,983 1,491 785 426 226 20 7,473 3,736 1,916 1,083 530 21 3,626 1,813 906 489 237 22 2,560 1,347 691 412 264 23 2,039 1,019 536 323 206 24 2,563 1,281 657 337 178 Those above are times in seconds for the 24 positions on 1,2,4,8,16 cpus. pos 1 2 4 8 16 1 87,735,974 89,052,012 105,025,123 109,467,495 155,514,410 2 88,954,757 90,289,077 100,568,301 110,988,161 137,965,406 3 101,302,792 102,822,332 111,433,074 117,366,515 119,271,093 4 71,726,853 72,802,754 74,853,409 88,137,085 104,230,094 5 49,386,616 50,127,414 55,834,316 61,619,298 89,506,306 6 133,238,718 135,237,296 146,562,594 168,838,428 226,225,307 7 58,593,747 62,602,792 66,243,490 68,868,878 93,575,946 8 225,906,282 229,294,872 248,496,917 261,728,552 340,548,431 9 122,264,617 124,098,584 138,226,951 159,930,005 199,204,874 10 75,301,353 76,430,872 80,651,716 83,656,702 93,431,597 11 95,321,494 96,751,315 104,853,646 107,369,070 123,994,812 12 79,975,416 85,447,418 85,657,884 94,000,085 112,174,209 13 66,100,160 70,622,802 70,796,754 78,834,155 96,053,649 14 58,099,574 58,971,066 67,561,507 74,791,668 95,627,150 15 84,143,340 85,405,488 92,557,676 97,486,065 124,516,703 16 75,738,094 80,920,173 79,039,499 84,141,904 94,701,972 17 154,901,225 184,970,278 242,480,013 279,166,418 416,426,105 18 20,266,629 22,856,254 28,443,165 31,608,146 42,454,639 19 93,858,903 95,266,785 100,527,830 108,742,238 114,692,731 20 231,206,390 234,674,482 241,284,621 271,751,263 264,493,531 21 112,457,464 114,144,324 114,425,474 123,247,294 118,558,091 22 81,302,340 86,865,131 89,432,576 106,348,704 135,196,568 23 63,598,940 64,552,923 68,117,815 81,871,010 103,621,303 24 80,413,971 81,620,179 83,919,196 85,810,169 90,074,814 >So despite 8 processors more, you didn't get twice as much as nodes >a second more. In fact you lost 60% speed there in nodes a second. > >This was all tested at the same machine you write in the same article. > >They were busy idling at the 16 processor? You have my _real_ data above as was published in the JICCA. I don't know what data you are quoting. But, for reference, on the first position in the test, the NPS is 31K for 1 processor, 62K for 2, 126K for 4, 251K for 8 and 500K for 16. I have no idea what numbers _you_ are talking about. The above tables and the NPS for the first position are simply taken from the JICCA DTS paper directly. So I have no way of answering your last question since it makes no sense. I can't find the numbers you gave Aha. I see. Please try again. You screwed up in copying the data down so your 8 processor number was too high because you picked the _wrong_ number of the 8 processor time. And you like to criticize me for errors? :) > >BTW, where are your 32 processor results? I have never seen them in any >publication. yet your test is pretty simple. Just a 24 positions and >you only need to get near the same depth of the 16 processor search. > I never ran that test. The 32 processor machine was _far_ faster than the 16 processor machine, which meant I would have had to re-run the 1,2,4,8 and 16 cpu numbers in addition to the 32 processor numbers. I never had access to a T90 long enough to do that. I did run a couple of the positions, and the curve was definitely getting "worse". One test position I have ran 20X faster. Another was 18X faster. Which is not great. But then again, 11X out of 16 is also not "great" but it is satisfactory... >Should be theoretical under an hour... > Nope. T90 CPU is far faster than a C90 CPU. _all_ results would have to be re-run. > > >Best regards, >Vincent > > > > >>those two programs get to the 1-2 second mark, the cpus stay busy, period. >>Just like Crafty. I don't believe DB had any such problem either, once >>you get beyond the first search result they outputted for a iteration... >> >>You could test your hypothesis however, by seeing if the first couple >>of branching factors are worse than the last couple for each search... >>If they are similar, this isn't a problem... >> >>I might try this tonight late, for one log file...
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.