Arasan 10.0 Intel JA builds

[Jim Ablett]

Arasan 10.0 by Jon Dart .
http://www.arasanchess.org/
Win32 & 64-bit Intel compiler 10 p.g.o builds

First 64-bit compile with my newly-installed Intel 64-bit
environment.
Modest speed increases with the 32-bit compiles, but
the 64-bit is a monster!

http://mihd.net/cfzd63
Mirrors:
http://www.zshare.net/download/52956774b02793/
http://FastFreeFileHosting.com/file/120 ... a-zip.html

Jim.

[Tony Thomas]

How much is modest Jim?

[Werner]

Hi Jim, thank you very much!
Now you can try to compile Glaurung 2.0.1 or 2.0.1 too

I made a speed test on my dual with single 64bit versions:

Engine: Arasan 10.0 (SMP) ori 64 bit (256 MB)
von Jon Dart

13.20 0:35 +0.18 1.Sf3 Sf6 2.Sc3 Sc6 3.d4 e6 4.e3 Lb4 5.Ld2 d5 6.Ld3 Dd6 7.O-O (29.979.008) 833
14.20 2:03 +0.06 1.Sf3 Sf6 2.Sc3 Sc6 3.d4 e6 4.e3 Lb4 5.a3 Le7 6.Le2 d5 7.O-O O-O (103.696.085) 841
15.03 3:24 +0.12 1.Sf3 Sf6 2.Sc3 Sc6 3.d4 e6 4.e3 d5 5.Lb5 Le7 6.O-O Dd6 7.g3 a6 8.Le2 (167.618.467) 819
15.03 3:24 +0.12 1.Sf3 Sf6 2.Sc3 Sc6 3.d4 e6 4.e3 d5 5.Lb5 Le7 6.O-O Dd6 7.g3 a6 8.Le2 (167.618.467) 819
Bester Zug: Sg1-f3 Zeit: 3:24.453 min K/s: 819.890 Knoten: 167.618.467

and:

Engine: Arasan 10.0 (SMP)jax64 (256 MB)
von Jon Dart

13.20 0:34 +0.18 1.Sf3 Sf6 2.Sc3 Sc6 3.d4 e6 4.e3 Lb4 5.Ld2 d5 6.Ld3 Dd6 7.O-O (30.663.948) 896
14.20 1:45 +0.06 1.Sf3 Sf6 2.Sc3 Sc6 3.d4 e6 4.e3 Lb4 5.a3 Le7 6.Le2 d5 7.O-O O-O (94.004.574) 888
15.08 3:24 +0.12 1.Sf3 Sf6 2.Sc3 Sc6 3.d4 e6 4.e3 d5 5.Lb5 Le7 6.O-O Dd6 7.g3 a6 8.Le2 (179.787.637) 879
15.08 3:24 +0.12 1.Sf3 Sf6 2.Sc3 Sc6 3.d4 e6 4.e3 d5 5.Lb5 Le7 6.O-O Dd6 7.g3 a6 8.Le2 (179.787.637) 879
Bester Zug: Sg1-f3 Zeit: 3:24.484 min K/s: 879.286 Knoten: 179.787.637

on 14/20 e.g. 888kns via 841 kns or in time 1:45 via 2:03

[Jim Ablett]

How much is modest Jim?

Arasan 10.0 32-bit JA build

Code: Select all

"BS2830-01" bm Qxf3
0	0.00	Qxa6	 +0.31	19
0	0.00	Qf5	 +0.31	41
1	0.01	Qf5	 +0.31	72
1	0.01	Qxa6	 +0.31	118
2	0.03	Qxa6	 -0.06	250
3	0.03	Qxa6	 +0.00	816
4	0.03	Qxa6	 -0.06	2427
4	0.05	Qf5	 -0.06	6820
5	0.06	Qf5	 +0.06	10920
6	0.17	Qf5	 +0.06	66832
7	0.39	Qf5	 +0.00	173966
7	0.72	Qxa6	 +0.00	338995
8	0.78	Qxa6	 +0.31	373240
9	1.70	Qxa6	 +0.31	834078
10	2.30	Qxa6	 +0.56	1189445
538.09k nodes/second.
1018200 regular nodes, 2237215 quiescence nodes.
1017183 searches of hash table, 425798 successful (41 percent).
582729 hash entries inserted, 234923 entries replaced, 5657 inserts failed.
null cutoffs: 369339 (11%)
pruning: 2891378 moves (futility), 73191 nodes (razored)
extensions: 183789 check, 14396 recapture, 841 pawn, 
            11333 threat, 11598 forced move, 0 repetition.
0 tablebase probes, 0 tablebase hits
0 splits, average thread usage=1.00
position hash hit rate = 12.01%
pawn hash hit rate     = 97.47%
king/pawn hash hit rate= 89.52%
king cover hash hit rate= 95.87%
lazy cutoff in eval    = 34.06%

Arasan 10.0 32-bit Jon Dart build

Code: Select all

"BS2830-01" bm Qxf3
0	0.00	Qxa6	 +0.31	19
0	0.00	Qf5	 +0.31	41
1	0.00	Qf5	 +0.31	72
1	0.00	Qxa6	 +0.31	118
2	0.00	Qxa6	 -0.06	250
3	0.00	Qxa6	 +0.00	812
4	0.02	Qxa6	 -0.06	2412
4	0.02	Qf5	 -0.06	6796
5	0.03	Qf5	 +0.06	10784
6	0.16	Qf5	 +0.06	66407
7	0.39	Qf5	 +0.00	172509
7	0.72	Qxa6	 +0.00	330470
8	0.80	Qxa6	 +0.31	363715
9	1.59	Qxa6	 +0.31	756353
10	2.91	Qxa6	 +0.56	1401285
510.73k nodes/second.
973891 regular nodes, 2116016 quiescence nodes.
972860 searches of hash table, 408037 successful (41 percent).
558905 hash entries inserted, 226327 entries replaced, 5177 inserts failed.
null cutoffs: 359333 (11%)
pruning: 2768693 moves (futility), 68781 nodes (razored)
extensions: 176436 check, 13647 recapture, 941 pawn, 
            11082 threat, 10525 forced move, 0 repetition.
0 tablebase probes, 0 tablebase hits
0 splits, average thread usage=1.00
position hash hit rate = 12.10%
pawn hash hit rate     = 97.38%
king/pawn hash hit rate= 89.02%
king cover hash hit rate= 95.76%
lazy cutoff in eval    = 33.28%

Jim.

[jdart]

Shouldn't you be getting the same node counts at each depth for the two builds (since this is not running multi-threaded)? If they are different then they are not behaving the same.

--Jon

[Jim Ablett]

Hi Jon,

Shouldn't you be getting the same node counts at each depth for the two builds (since this is not running multi-threaded)?
If they are different then they are not behaving the same.

One of the Intel compiler optimization switches that I use is '/Qparallel' which auto-parallelizes loops in the code.
Maybe this affects epd test causing different nodecounts at each depth.

Jim.

[Jim Ablett]

Hi Werner,

Now you can try to compile Glaurung 2.0.1 or 2.0.1 too

Glaurung 64-bit JA build available on my homepage now.

The Glaurung 64-bit compile Brian posted earlier is very fast though.
Not much difference between the two on my system.

Jim.

[jdart]

One of the Intel compiler optimization switches that I use is '/Qparallel' which auto-parallelizes loops in the code

I'm not sure that's a good idea. You'll get better performance on a dual-core or dual-processor machine, but running the engine with two search threads would be even better. On the other hand, if you are running something like an engine match, you'd generally want each engine to use only a limited number of threads (1, on a dual-core box), and /Qparallel defeats that.

[Tony Thomas]

So, you are talking about a speed up of 4.5 to 5 %? I am trying to make sure that my calculations are correct.

[Jim Ablett]

On the other hand, if you are running something like an engine match, you'd generally want each engine to use only a limited number of threads (1, on a dual-core box), and /Qparallel defeats that.

I was under the impression with using /Qparallel switch that 2 thread loops would only be utilized if >

1. At least dual processors were available.
2. Engine was set to run on at least 2 threads.

I'll run some tests to analyse.

Jim.