Author: Vincent Diepeveen
Date: 08:49:28 04/13/03
Go up one level in this thread
On April 13, 2003 at 11:27:53, Robert Hyatt wrote: I said initially. It drops back to 10 splits a second in DIEP after a while. Search depth matters. Let's compare 2 things. time=45.98 cpu=464% mat=0 n=37870294 fh=88% nps=823k ext-> chk=638414 cap=249442 pp=9588 1rep=32966 mate=223 predicted=0 nodes=37870294 evals=14565859 endgame tablebase-> probes done=0 successful=0 hashing-> trans/ref=28% pawn=93% used=28% SMP-> split=431 stop=57 data=6/64 cpu=3:33 elap=45.98 MT 2 crafty 18.10 which i have here. 431 splits at 45 seconds. I guess you must limit in crafty the number of splits a lot as splitting is expensive in crafty when compared to the costs of a single node. Let's ignore the cpu=464% i do not understand why it says that. I have it at mt=2. probably small i/o bug. Now let's diep search for around this time: Took 0.12 seconds to start all 1 other processes out of 2 00:00 21 0k 0 0 21 (2) 2 (0,0) -0.022 Ng1-f3 d7-d5 ++ d2-d4 procnr=0 terug=1 org=[-22;-21] newwindow=[-22;520000] 00:00 71 0k 0 0 71 (2) 2 (0,0) 0.001 d2-d4 d7-d5 00:00 175 0k 0 0 175 (2) 3 (0,2) 0.157 d2-d4 d7-d5 Ng1-f3 00:00 443 0k 0 0 443 (2) 4 (0,5) 0.001 d2-d4 d7-d5 Ng1-f3 Ng8-f6 00:00 150800 151k 0 0 1508 (2) 5 (0,19) 0.190 d2-d4 d7-d5 Ng1-f3 Ng8-f6 Nb1-c3 00:00 318900 319k 0 0 3189 (2) 6 (0,27) 0.001 d2-d4 d7-d5 Ng1-f3 Ng8-f6 Nb1-c3 N b8-c6 00:00 149744 150k 0 0 13477 (2) 7 (3,68) 0.179 d2-d4 d7-d5 Ng1-f3 Ng8-f6 Bc1-f4 Nf6-h5 Bf4-g5 00:00 136110 136k 0 0 27222 (2) 8 (6,147) 0.001 d2-d4 d7-d5 Ng1-f3 Ng8-f6 Bc1-f4 Nf6-h5 Bf4-g5 Nb8-c6 00:01 127109 127k 0 0 205917 (2) 9 (45,502) 0.105 d2-d4 Ng8-f6 Nb1-c3 Nb8-c6 Bc1 -f4 d7-d6 Ng1-f3 Bc8-f5 e2-e3 00:04 127013 127k 0 0 572829 (2) 10 (76,666) 0.001 d2-d4 Ng8-f6 Nb1-c3 d7-d5 Bc1 -f4 Bc8-f5 Ng1-f3 Nb8-c6 Nf3-e5 Nf6-e4 00:17 152655 153k 0 0 2648566 (2) 11 (330,1980) 0.108 d2-d4 d7-d5 Ng1-f3 Nb8-c6 Nb1-c3 Bc8-f5 Nf3-h4 Bf5-c8 Bc1-g5 Ng8-f6 e2-e3 00:38 154041 154k 0 0 5889009 (2) 12 (743,4189) 0.008 d2-d4 d7-d5 Bc1-f4 Bc8-f5 Ng1-f3 Ng8-f6 Nb1-c3 Nb8-c6 Nc3-b5 Ra8-c8 Nf3-e5 Nc6xe5 d4xe5 Of course if i use same conditions like crafty when to split then it will look different with regards to the number of splits performed. Splitting in diep is very cheap. I already split >= 2 ply left searches and i split quickly in current versions. The reason is that you get 500 cpu's quicker busy and find bugs sooner. No doubt in future i will again optimize it to a state where it will optimize search depth more at x86. If that's with many splits a second at 2-4 processes, then i'll go for that. If it is with less splits a second i'll go for that. Note that the 4189 number at 12 ply is not the number of splits only, it is the total number of searches. So about 11*20 + 1 = 220 + 1 = 221 are from searching the root. >On April 13, 2003 at 08:32:37, Vincent Diepeveen wrote: > >>On April 13, 2003 at 08:21:42, Vincent Diepeveen wrote: >> >>>On April 13, 2003 at 02:37:57, Tom Kerrigan wrote: >>> >>>>On April 13, 2003 at 01:04:52, Robert Hyatt wrote: >>>> >>>>>It _is_ pinned on SMT. The two logical processors are producing wildly >>>>>imbalanced results when using threads, vs using two separate processes. It >>>>>would appear to be cache-related... >>>> >>>>This is some sort of joke, right? You and Vincent see the same behavior, you >>>>have SMT and Vincent doesn't, and somehow the problem is with SMT? >>>> >>>>How much of the time are your threads idle, out of curiosity? If one thread is >>>>idle much more than the other, then of course that is going to skew your NPS. >>>> >>>>-Tom >>> >>>Of course both Crafty and DIEP are using YBW. I didn't checkout what bob does >>>here, but in past in DIEP i used to always let process 0 let the search start. >>>Nowadays that is not the case. The i/o thread picks the first process it can >>>get. All search processes are completely identical. This process then is >>>starting the search. That means the other CPUs idle when this process starts the >>>search. >> >>also read that 'idle' not in litterary sense. Letting them REALLY idle with >>sleep() or WaitForSingleObject, is at a REAL smp system (like dual K7) just too >>expensive. Latency to wake up processors is at sick high levels. 15 ms just like >>that. Imagine that because of the YBW search, you have to split initially like >>50-100 times a second. 15ms is death sentence. So 'idle' cpu's are spinning >>around until at a shared memory variable some flag is set. I let them do some >>arithmetic function for a 100 times while 'idling'. > >If you do this right you won't split _that_ often. > > time=35.97 cpu=381% mat=-1 n=80006982 fh=92% nps=2224k > ext-> chk=1487513 cap=353299 pp=32860 1rep=79236 mate=15135 > predicted=3 nodes=80006982 evals=19493470 > endgame tablebase-> probes done=0 successful=0 > SMP-> split=1840 stop=163 data=15/64 cpu=2:17 elap=35.97 > time used: 29.81 > > >In the above from a game on ICC, in 35 seconds, I did 1800 splits total. The >deeper the search the better this becomes... > > time=2:33 cpu=396% mat=0 n=282753699 fh=91% nps=1840k > ext-> chk=3046093 cap=1083298 pp=16735 1rep=192964 mate=3400 > predicted=8 nodes=282753699 evals=114936261 > endgame tablebase-> probes done=0 successful=0 > SMP-> split=2683 stop=424 data=15/64 cpu=10:09 elap=2:33 > time used: 8.29 > > time=4:03 cpu=396% mat=0 n=466004128 fh=90% nps=1911k > ext-> chk=3120074 cap=1773259 pp=60704 1rep=227466 mate=5595 > predicted=9 nodes=466004128 evals=160300467 > endgame tablebase-> probes done=0 successful=0 > SMP-> split=5811 stop=950 data=18/64 cpu=16:06 elap=4:03 > time used: 2:43 > > time=3:47 cpu=396% mat=0 n=421757405 fh=92% nps=1855k > ext-> chk=3436512 cap=1222511 pp=75583 1rep=186606 mate=3165 > predicted=12 nodes=421757405 evals=149496490 > endgame tablebase-> probes done=0 successful=0 > SMP-> split=3524 stop=337 data=17/64 cpu=15:01 elap=3:47 > >> >>>In crafty that's also the case, but i do not know whether Bob always picks a >>>certain thread as first. If so then that might explain quite something. >>> >>>Measuring idle time with SMT is very hard to do objective, but of course you can >>>relatively check it out. Basically the problem is you do not know what the >>>maximum % is that i can get out of SMT, because it is dependant upon the other >>>process too.
This page took 0.01 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.