230 likes | 237 Views
This summary highlights the work of a technical student at CERN LCG AA Meeting in September 2004, covering test and validation studies of mathematical software libraries. Topics include special functions, distributions, random numbers, frequency tests, point tests, numerical integration, and linear algebra. Results and conclusions are provided.
E N D
Test and Validation Studies of Mathematical Software Libraries A summary of my work as a technical student at CERN LCG AA Meeting, 22. September 2004
Special Functions • Comparison of numerical results • Performance • GSL-NAG-C and GSL-TMath • Bessel functions • Gamma, Logarithm of Gamma, Error function and Complementary Error function
Results • GSL performed very well compared to NAG-C • Difference usually less than the estimated error • Bigger differences between GSL and TMath • Little difference in time for GSL • NAG-C and TMath are faster
Timing of Special Functions • 100 000 function calls for Bessel I0, I1, J0 and J1 • 50 000 function calls for the rest
Distributions • Comparison of numerical results • Performance in the evaluation • Generation of random numbers according to distribution • Comparison and Kolmogorov-Smirnov test • Normal distribution, Landau distribution, Gamma distribution, Poisson distribution and Chi-Square distribution
Timing Results for some Distributions • 1 000 000 function calls
Random Numbers • Two tests • Frequency test • Point test • Main generators from GSL • gsl_rng_mt19937 • gsl_rng_cmrg • gsl_rng_mrg • gsl_rng_taus • gsl_rng_taus2 • gsl_rng_gfsr4 • gsl_rng_ranlux389 • gsl_rng_ranlux • gsl_rng_ranlxd2
Frequency Test Fill space in d dimensions with points formed from a sequence of random numbers. Look in a small volume and the frequency as the number of bins which maximize |Nodd-Neven|. gsl_rng_minstd
Frequency Test • With this frequency, look other places in the space and compute Nodd. • Nodd should be normal distributed
Results • 10 results for Nodd • Kolmogorov-Smirnov test • Taus, 8 dim and Ranlux389, 6 dim • New test for poor results • All passed
Point Test • Arrange a sequence of random numbers into multidimensional points • Define distance between two points as • Find all points Pi that are closer to P1 than the mean-n*standard deviation (n=3,4,5) • Calculate the distance between Pi+1 and P2
Point test • For the distance should be normal distributed. • Use Kolmogorov-Smirnov test • All generators pass
Numerical Integration • Wrapper for existing gsl algorithms • Tested on a few number of integrals • Compare numerical results with analytical results • No difference larger than 10-7 (input tolerance), but need further testing
Quadrature routines QAG – adaptive integration QAGUI – adaptive integration from zero to infinity QAGS – adaptive integration with singularities QNG – non-adaptive Gauss-Kronrod integration NB! Different integrals are used, marked with (*) on last slide Performance in Numerical Integration
Linear Algebra • E. Myklebust summer student 2003 • A Comparative study of Numerical Linear Algebra Libraries • Particle Track Reconstruction, Kalman filter update equations • Multiplication, addition, inversion and transpose • Originally 2x2, 2x5, 5x2 and 5x5 • Extended to bigger matrices • 2x5, 4x10, 10x25 and 20x50 • CLHEP, uBLAS, LAPACK, GSL and ROOT • Used timer from SEAL base
Results (Linux, P4 1.8 MHz ) • High RMS for GSL and LAPACK in 2x5 • Error with ROOT for 20x50
Conclusions • GSL performs reasonably good • Both tests of randomness were passed by all the main generators from GSL • More testing is needed for the numerical integration • All test programs are in the SEAL cvs repository • A test suite can be easily created and automatically run for new SEAL releases • A written report of my work will be put on the webpage when finished