SDPA クラスタでの SDPARA の実験 その３

図に示したブロック対角構造を持った非常に大きな SDP を SDPA クラスタで解いた結果を示す（もちろん SDPARA を用いる）。参考までに以前 AIST Super Cluster (ASC) M-64 で解いた結果を始めに示す。

問題名 : H2O.1A1.DZ.pqgt1t2p.dat-s

1: ASC M-64 クラスタ
CPU : Itanium 2 1.3GHz
メモリ : 16GB

　 8 CPU を用いて実行時間は 2060237.619 sec. = およそ 24 日

phase.value = pdFEAS
Iteration = 43
mu = 8.0967296259055195e-10
relative gap = 5.2793022584586229e-08
gap = 1.2885135526666044e-05
digits = 7.2774234724141191e+00
objValPrimal = -8.5344849104e+01
objValDual = -8.5344853610e+01
p.feas.error = 3.1098588468e-08
d.feas.error = 6.3761717684e-08

Time(sec) Ratio(% : MainLoop)
Predictor time = 2037533.084541, 99.068638
Corrector time = 11533.887862, 0.560799
Make bMat time = 1985337.814099, 96.530807
Make bDia time = 8273.927409, 0.402294
Make bF1 time = 105.914069, 0.005150
Make bF2 time = 17018.234065, 0.827458
Make bF3 time = 1959912.087295, 95.294561
Make bPRE time = 57.191760, 0.002781
Make rMat time = 2627.625670, 0.127760
Make gVec Mul = 7118.398276, 0.346110
Make gVec time = 11.590774, 0.000564
Copy bMat time = 18839.678796, 0.916020
Symm bMat time = 815.432378, 0.039648
Cholesky bMat = 22137.972316, 1.076389
solve = 54.271096, 0.002639
Copy DyVec time = 0.266431, 0.000013
Ste Pre time = 0.012160, 0.000001
Ste Cor time = 702.919505, 0.034177
sumDz = 136.354752, 0.006630
makedX = 10381.292192, 0.504758
symmetriseDx = 72.610232, 0.003530
makedXdZ = 10604.559229, 0.515613
xMatTime = 2123.026743, 0.103225
zMatTime = 2122.108908, 0.103181
invzMatTime = 1377.336452, 0.066969
xMatzMatTime = 2535.852326, 0.123298
EigxMatTime = 364.493395, 0.017722
EigzMatTime = 325.059380, 0.015805
EigxMatzMatTime = 0.000001, 0.000000
updateRes = 122.320466, 0.005947
EigTime = 689.552776, 0.033527
sub_total_bMat = 71350.486096, 3.469193
Parallel = 2027197.025890, 98.566080
Main Loop = 2056688.300195, 100.000000
File Check = 0.000000, 0.000000
File Change = 0.096272, 0.000005
File Trans = 19.140806, 0.000931
File Read = 64.160690, 0.003120
Total = 2060237.618669, 100.172574