Author: Eugene Nalimov
Date: 14:07:21 12/17/02
Go up one level in this thread
That's pointless. If somebody does not work for him (or if he just thinks it cannot work for him), it should not work for anymbody. BTW, I have simple explanation why HT does not work very well for Diep. HT means "hyper threading", not "hyper processing". When you have different processes, you actually halved effective size of L1/L2 data cache. If your program accesses several copies of the data -- be it bitmap masks, piece/square tables, bit count tables, piece attack tables, etc. -- when you have several processes you have to duplicate that data in the cache. Situation with I-cache is better, at least on Windows. If you have several copies of the same executable running, and those copied were loaded from the same address (no relocation was necessary), Windows maps all those executables to the same physical memory. That means that if different processess execute the same code it should be in the trace cache (if it fit into 12k). And there are much more interesting and subtle issues with shared memory. Very annoying things can happen if you have 2 different regions of physical memory mapped to the same addresses of the virtual memory in 2 threads running on the same physical CPU... Of course that's all waiving hands, but I believe I can waive my hands not worse than someone else. Thanks, Eugene On December 17, 2002 at 16:45:37, Robert Hyatt wrote: >On December 17, 2002 at 11:46:57, Vincent Diepeveen wrote: > >>On December 17, 2002 at 11:29:19, Robert Hyatt wrote: >> >>hello Bob, >> >>please do the same tests i did with DIEP too with crafty. >> >>Of course as you always say that doing a few tests proof nothing, >>please repeat them twice. >> >>For me doing a test twice with crafty is sufficient. > >I generally say "your doing a few tests proves nothing." "Why?" you ask? >Because >you seem incapable of understanding simple ideas. Remember your nonsense about >"on a dual Crafty runs _no_ faster at all here..."??? So if I remain a bit >suspicious of >any number(s) you report, there is a reason for it... > > >I ran 24 positions twice and reported the NPS for 1 thread, no SMT, two threads, >no SMT, >three threads, SMT on, and four threads, SMT on. > >What more can I run??? > > >> >>I am especially interested in the completed logs too so that we all can >>see what mainline you took to compare the speedup and absolute speeds >>in nps. > >All I reported was NPS. I'm not going to post such a huge wad of output here. >I will >take one position from each of the four tests and give you those. I have no >idea what you >mean by "what mainline you took to compare the speedup and absolute speeds in >NPS." >I didn't take _any_ mainline. I didn't report _any_ speedup. I only reported >the increase in >raw NPS numbers. So the rest of your query simply makes no sense to me. This >is not about >parallel search efficiency. It is about whether SMT speeds things up or not, >and the answer (so >far) is clearly "yes it does." > >This is the last position from the 24 I ran. It is one of the Kopec positions >but It >doesn't say which one. The position is this (FEN): >3rn2k/ppb2rpp/2ppqp2/5N2/2P1P3/1P5Q/PB3PPP/3RR1K1 w > >Run 1. one thread, no SMT: >White(1): move > clearing hash tables > time surplus 0.00 time limit 166:39 (166:39) > depth time score variation (1) > 1 0.00 0.81 1. Bd4 > 1-> 0.00 0.81 1. Bd4 > 2 0.00 0.73 1. Bd4 Bb6 > 2-> 0.00 0.73 1. Bd4 Bb6 > 3 0.00 -- 1. Bd4 > 3 0.00 0.29 1. Bd4 g6 2. Nh6 Qxh3 3. gxh3 > 3 0.01 0.58 1. Nh6 Re7 2. Qxe6 Rxe6 > 3 0.01 0.63 1. Qf3 g6 2. Nh6 > 3-> 0.01 0.63 1. Qf3 g6 2. Nh6 > 4 0.01 0.53 1. Qf3 g6 2. Nh6 Re7 > 4 0.01 0.58 1. Nh6 Re7 2. Qxe6 Rxe6 > 4-> 0.04 0.58 1. Nh6 Re7 2. Qxe6 Rxe6 > 5 0.04 ++ 1. Nh6!! > 5 0.05 2.19 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 5-> 0.06 2.19 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 6 0.06 2.40 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 > 6-> 0.10 2.40 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 > 7 0.11 2.32 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 Bb6 > 7-> 0.16 2.32 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 Bb6 > 8 0.20 2.31 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 b5 5. cxb5 cxb5 > 8-> 0.33 2.31 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 b5 5. cxb5 cxb5 > 9 0.37 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 Ke6 > 9-> 0.54 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 Ke6 > 10 0.61 2.41 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 b5 6. cxb5 cxb5 > 10-> 1.49 2.41 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 b5 6. cxb5 cxb5 > 11 1.62 2.37 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Rg1 d5 7. > cxd5 cxd5 > 11-> 3.93 2.37 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Rg1 d5 7. > cxd5 cxd5 > 12 4.22 2.46 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Rd7 5. Kf3 Re7 6. a4 Bb6 7. > Rg1 > 12-> 7.90 2.46 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Rd7 5. Kf3 Re7 6. a4 Bb6 7. > Rg1 > 13 8.84 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Ba3 Ke7 7. > Bb4 Ne6 > 13-> 25.57 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Ba3 Ke7 7. > Bb4 Ne6 > time=25.57 cpu=99% mat=0 n=31129467 fh=94% nps=1217k > ext-> chk=917605 cap=59244 pp=3275 1rep=119119 mate=34241 > predicted=0 nodes=31129467 evals=3804978 > endgame tablebase-> probes done=0 successful=0 > SMP-> split=0 stop=0 data=0/64 cpu=25.50 elap=25.57 > >Run 2: two threads, no SMT: > > depth time score variation (1) >starting thread 1 > 1 0.00 0.81 1. Bd4 > 1-> 0.00 0.81 1. Bd4 > 2 0.00 0.73 1. Bd4 Bb6 > 2-> 0.00 0.73 1. Bd4 Bb6 > 3 0.00 -- 1. Bd4 > 3 0.00 0.29 1. Bd4 g6 2. Nh6 Qxh3 3. gxh3 > 3 0.00 0.58 1. Nh6 Re7 2. Qxe6 Rxe6 > 3 0.01 0.63 1. Qf3 g6 2. Nh6 > 3-> 0.07 0.63 1. Qf3 g6 2. Nh6 > 4 0.07 0.53 1. Qf3 g6 2. Nh6 Re7 > 4 0.07 0.58 1. Nh6 Re7 2. Qxe6 Rxe6 > 4-> 0.10 0.58 1. Nh6 Re7 2. Qxe6 Rxe6 > 5 0.10 ++ 1. Nh6!! > 5 0.11 2.19 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 5-> 0.14 2.19 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 6 0.15 2.40 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 > 6-> 0.17 2.40 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 (s=4) > 7 0.18 2.32 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 Bb6 (s=3) > 7-> 0.35 2.32 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 Bb6 (s=6) > 8 0.37 2.31 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 b5 5. cxb5 cxb5 (s=5) > 8-> 0.50 2.31 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 b5 5. cxb5 cxb5 (s=4) > 9 0.57 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 Ke6 (s=3) > 9-> 0.73 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 Ke6 (s=5) > 10 0.78 2.41 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 b5 6. cxb5 cxb5 (s=4) > 10-> 1.38 2.41 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 b5 6. cxb5 cxb5 (s=8) > 11 1.49 2.37 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Rg1 d5 7. > cxd5 cxd5 (s=7) > 11-> 2.98 2.37 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Rg1 d5 7. > cxd5 cxd5 (s=10) > 12 3.19 2.46 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Rd7 5. Kf3 Re7 6. a4 Bb6 7. > Rg1 (s=9) > 12-> 6.02 2.46 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Rd7 5. Kf3 Re7 6. a4 Bb6 7. > Rg1 (s=9) > 13 6.70 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Ba3 Ke7 7. > Bb4 Ne6 (s=8) > 13-> 18.42 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Ba3 Ke7 7. > Bb4 Ne6 (s=12) > time=18.42 cpu=188% mat=0 n=34809013 fh=94% nps=1889k > ext-> chk=992883 cap=70671 pp=5213 1rep=135350 mate=27885 > predicted=0 nodes=34809013 evals=4659923 > endgame tablebase-> probes done=0 successful=0 > SMP-> split=761 stop=38 data=7/64 cpu=34.74 elap=18.42 > >Run three: three threads, SMT _on_: > >White(1): move > clearing hash tables > time surplus 0.00 time limit 166:39 (166:39) > depth time score variation (1) >starting thread 1 >starting thread 2 > 1 0.00 0.81 1. Bd4 > 1-> 0.01 0.81 1. Bd4 > 2 0.01 0.73 1. Bd4 Bb6 > 2-> 0.01 0.73 1. Bd4 Bb6 > 3 0.01 -- 1. Bd4 > 3 0.01 0.29 1. Bd4 g6 2. Nh6 Qxh3 3. gxh3 > 3 0.01 0.58 1. Nh6 Re7 2. Qxe6 Rxe6 > 3 0.02 0.63 1. Qf3 g6 2. Nh6 > 3-> 0.08 0.63 1. Qf3 g6 2. Nh6 > 4 0.08 0.53 1. Qf3 g6 2. Nh6 Re7 > 4 0.09 0.58 1. Nh6 Re7 2. Qxe6 Rxe6 > 4-> 0.10 0.58 1. Nh6 Re7 2. Qxe6 Rxe6 > 5 0.11 ++ 1. Nh6!! > 5 0.11 2.19 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 5-> 0.16 2.19 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 6 0.17 2.40 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 > 6-> 0.19 2.40 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 (s=4) > 7 0.20 2.32 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 Bb6 (s=3) > 7-> 0.24 2.32 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 Bb6 (s=7) > 8 0.26 2.31 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 b5 5. cxb5 cxb5 (s=6) > 8-> 0.36 2.31 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 b5 5. cxb5 cxb5 (s=4) > 9 0.38 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 Ke6 (s=3) > 9-> 0.50 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 Ke6 (s=5) > 10 0.54 2.41 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 b5 6. cxb5 cxb5 (s=4) > 10-> 1.07 2.41 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 b5 6. cxb5 cxb5 (s=8) > 11 1.16 2.37 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Rg1 d5 7. > cxd5 cxd5 (s=7) > 11-> 2.66 2.37 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Rg1 d5 7. > cxd5 cxd5 (s=10) > 12 2.86 2.46 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Rd7 5. Kf3 Re7 6. a4 Bb6 7. > Rg1 (s=9) > 12-> 5.40 2.46 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Rd7 5. Kf3 Re7 6. a4 Bb6 7. > Rg1 (s=8) > 13 6.02 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Ba3 Ke7 7. > Bb4 Ne6 (s=7) > 13-> 17.69 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Ba3 Ke7 7. > Bb4 Ne6 (s=13) > time=17.69 cpu=296% mat=0 n=38066840 fh=94% nps=2151k > ext-> chk=1076447 cap=74646 pp=5437 1rep=149438 mate=35093 > predicted=0 nodes=38066840 evals=5253700 > endgame tablebase-> probes done=0 successful=0 > SMP-> split=2994 stop=217 data=15/64 cpu=52.47 elap=17.69 > >Run four: four threads, SMT _on_: > >White(1): move > clearing hash tables > time surplus 0.00 time limit 166:39 (166:39) > depth time score variation (1) >starting thread 1 >starting thread 2 >starting thread 3 > 1 0.00 0.81 1. Bd4 > 1-> 0.03 0.81 1. Bd4 > 2 0.06 0.73 1. Bd4 Bb6 > 2-> 0.10 0.73 1. Bd4 Bb6 > 3 0.10 -- 1. Bd4 > 3 0.11 0.29 1. Bd4 g6 2. Nh6 Qxh3 3. gxh3 > 3 0.11 0.58 1. Nh6 Re7 2. Qxe6 Rxe6 > 3 0.11 0.63 1. Qf3 g6 2. Nh6 > 3-> 0.14 0.63 1. Qf3 g6 2. Nh6 > 4 0.14 0.53 1. Qf3 g6 2. Nh6 Re7 > 4 0.14 0.58 1. Nh6 Re7 2. Qxe6 Rxe6 > 4-> 0.15 0.58 1. Nh6 Re7 2. Qxe6 Rxe6 > 5 0.16 ++ 1. Nh6!! > 5 0.19 2.19 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 5-> 0.20 2.19 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 6 0.20 2.40 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 > 6-> 0.28 2.40 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 (s=4) > 7 0.35 2.32 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 Bb6 (s=3) > 7-> 0.43 2.32 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Bd4 Bb6 (s=7) > 8 0.45 2.31 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 b5 5. cxb5 cxb5 (s=6) > 8-> 0.60 2.31 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 b5 5. cxb5 cxb5 (s=4) > 9 0.65 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 Ke6 (s=3) > 9-> 0.76 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 Ke6 (s=5) > 10 0.83 2.41 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 b5 6. cxb5 cxb5 (s=4) > 10-> 1.36 2.41 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Ba5 5. Rg1 b5 6. cxb5 cxb5 (s=8) > 11 1.46 2.37 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Rg1 d5 7. > cxd5 cxd5 (s=7) > 11-> 3.06 2.37 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Rg1 d5 7. > cxd5 cxd5 (s=10) > 12 3.27 2.46 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Rd7 5. Kf3 Re7 6. a4 Bb6 7. > Rg1 (s=9) > 12-> 5.66 2.46 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Rd7 5. Kf3 Re7 6. a4 Bb6 7. > Rg1 (s=8) > 13 6.34 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Ba3 Ke7 7. > Bb4 Ne6 (s=7) > 13-> 16.09 2.34 1. Nh6 Qxh3 2. Nxf7+ Kg8 3. gxh3 Kxf7 > 4. Kg2 Bb6 5. Kf3 Nc7 6. Ba3 Ke7 7. > Bb4 Ne6 (s=13) > time=16.09 cpu=377% mat=0 n=36609790 fh=94% nps=2275k > ext-> chk=1035481 cap=72332 pp=4034 1rep=144147 mate=40133 > predicted=0 nodes=36609790 evals=4935572 > endgame tablebase-> probes done=0 successful=0 > SMP-> split=5501 stop=394 data=16/64 cpu=1:00 elap=16.09 > > > > >>>>>parallel search overhead, you have a problem on _normal_ SMP machines as well. >>>> >>>>Indeed it is true that the first seconds the HT/SMT gives big problems >>>>in speed. Only after a couple of minutes the speed shows. I see only >>>>a speedup after a minute or 3 each position. >>> >>>So? That is _your_ program's results. Mine are just like they have always >>>been. I get a reasonable speedup whether it is one second per move or one >>>hour per move. No difference. >>> >>> >>>> >>>>I need to add however that i could improve a few issues in this version >>>>which could get that down to 1 minute but like you i doubt whether the 11.4% >>>>of HT is worth it. >>> >>>11.4% is _always_ worth it IMHO... >>> >>> >>>> >>>>I prefer a dual AMD instead for the moment! >>>> >>> >>> >>>Nothing wrong with that. I got the dual xeon because I wanted a chance to play >>>with the SMT stuff since it is obviously going to be "the future" of >>>microprocessor >>>architecture...
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.