Author: Robert Hyatt
Date: 10:09:08 05/11/04
Since it appears that I am the _only_ person willing to post actual raw data and speedup summary data for a parallel search, here is the third installment. (#1 was Martin's analysis of the CB positions on the opteron). (#2 was my post on the same analysis but for the BK positions). This again uses the Cray Blitz positions from the DTS paper, but this time I ran them on my dual xeon. Note that this is the current version of Crafty rather than the identical version I used on the opteron, so scores can be a little different. First, the data is once again at ftp.cis.uab.edu/pub/hyatt/smpdata, but the filenames for these tests all end in "xeon". Here is the raw summary: pos T T2 T2 T2 T2 1 105 55/1.91 56/1.88 56/1.88 58/1.81 2 209 112/1.87 114/1.83 114/1.83 112/1.87 3 230 135/1.70 134/1.72 132/1.74 137/1.68 4 235 165/1.42 145/1.62 145/1.62 144/1.63 5 122 69/1.77 62/1.97 64/1.91 69/1.77 6 101 60/1.68 65/1.55 61/1.66 60/1.68 7 109 60/1.82 60/1.82 60/1.82 59/1.85 8 225 142/1.58 143/1.57 153/1.47 148/1.52 9 98 74/1.32 69/1.42 76/1.29 63/1.56 10 149 112/1.33 92/1.62 93/1.60 93/1.60 11 263 153/1.72 139/1.89 142/1.85 154/1.71 12 215 131/1.64 118/1.82 130/1.65 123/1.75 13 299 202/1.48 191/1.57 196/1.53 194/1.54 14 182 100/1.82 100/1.82 98/1.86 98/1.86 15 142 96/1.48 85/1.67 99/1.43 87/1.63 16 286 212/1.35 176/1.62 156/1.83 250/1.14 17 256 164/1.56 164/1.56 164/1.56 170/1.51 18 173 102/1.70 120/1.44 98/1.77 98/1.77 19 159 60/2.65 55/2.89 56/2.84 60/2.65 20 197 105/1.88 117/1.68 113/1.74 110/1.79 21 97 62/1.56 57/1.70 42/2.31 81/1.20 22 218 98/2.22 104/2.10 160/1.36 120/1.82 23 95 83/1.14 95/1.00 82/1.16 107/0.89 24 119 100/1.19 91/1.31 159/0.75 167/0.71 average SU 1.66 1.71 1.69 1.62 T is the 1-cpu time to last completed iteration. T2 is the time to complete the same iteration with 2 processors. The second number in the column is the speedup. Again 3 of 4 hit my 1.7X number right on, one rounds to 1.6X. I suppose Vincent can feel free to use that column to justify his "1.7X is wrong" of course. A couple of notes. The last 2-3 searches are not very good test positions. They are beginning to see score and PV changes at the same depth. As a result, a 2-cpu 14 ply search can actually return a better (more accurate) result than the 1-cpu test if you are lucky. But any actual PV score changes affect the size of the tree, and hence the speedup. In any case, now after 12 test runs with 2 cpus, 4 for CB positions on opteron, 4 for BK positions on opteron, 4 for CB positions on xeon, it would _appear_ to all but the most thickheaded observer (we know who that is) that the 1.7 estimate for my speedup approximation formula is pretty accurate. From 8 more tests we also saw that the 4-cpu estimate was also pretty accurate. Regardless of what you might read here. :) At some point I'll post some 8-cpu numbers as well.. to show how the 5.9X estimate pans out. I've already run the test, but until such machines are available to others, there's little point in starting yet another disinformation war until someone can repeat the test independently. BTW if anyone wants to run this test on their own dual or quad box, let me know and I can give you the test positions and crafty input file to run the tests...
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.