Lecture1

403 views

Published on

0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
403
On SlideShare
0
From Embeds
0
Number of Embeds
2
Actions
Shares
0
Downloads
5
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

Lecture1

  1. 1. Advanced Computational Methods in Statistics: Lecture 1 Monte Carlo Simulation & Introduction to Parallel Computing Axel Gandy Department of Mathematics Imperial College London http://www2.imperial.ac.uk/~agandy London Taught Course Centre for PhD Students in the Mathematical Sciences Spring 2011
  2. 2. Today’s LecturePart I Monte Carlo SimulationPart II Introduction to Parallel Computing Axel Gandy 2
  3. 3. Random Number Generation Computation of Integrals Variance Reduction Techniques Part I Monte Carlo Simulation Random Number Generation Computation of Integrals Variance Reduction Techniques Axel Gandy Monte Carlo Simulation 3
  4. 4. Random Number Generation Computation of Integrals Variance Reduction TechniquesUniform Random Number Generation Basic building block of simulation: stream of independent rv U1 , U2 , . . . ∼ U(0, 1) “True” random number generators: based on physical phenomena Example http://www.random.org/; R-package random: “The randomness comes from atmospheric noise” Disadvantages of physical systems: cumbersome to install and run costly slow cannot reproduce the exact same sequence twice [verification, debugging, comparing algorithms with the same stream] Pseudo Random Number Generators: Deterministic algorithms Example: linear congruential generators: sn un = , sn+1 = (asn + c)modM M Axel Gandy Monte Carlo Simulation 5
  5. 5. Random Number Generation Computation of Integrals Variance Reduction TechniquesGeneral framework for Uniform RNG (L’Ecuyer, 1994) T s1 T s2 T s3 T ... s G G G u1 u2 u3 s initial state (’seed’) S finite set of states T : S → S is the transition function U finite set of output symbols (often {0, . . . , m − 1} or a finite subset of [0, 1]) G : S → U output function si := T (Si−1 ) and ui := G (si ). output: u1 , u2 , . . . Axel Gandy Monte Carlo Simulation 6
  6. 6. Random Number Generation Computation of Integrals Variance Reduction TechniquesSome Notes for Uniform RNG S finite =⇒ ui is periodic In practice: seed s often chosen by clock time as default. Good practice to be able to reproduce simulations: Save the seed! Axel Gandy Monte Carlo Simulation 7
  7. 7. Random Number Generation Computation of Integrals Variance Reduction TechniquesQuality of Random Number Generators “Random numbers should not be generated with a method chosen at random” (Knuth, 1981, p.5) Some old implementations were unreliable! Desirable properties of random number generators: Statistical uniformity and unpredictability Period Length Efficiency Theoretical Support Repeatability, portability, jumping ahead, ease of implementation (more on this see e.g. Gentle (2003), L’Ecuyer (2004), L’Ecuyer (2006), Knuth (1998)) Usually you will do well with generators in modern software (e.g. the default generators in R). Don’t try to implement your own generator! (unless you have very good reasons) Axel Gandy Monte Carlo Simulation 8
  8. 8. Random Number Generation Computation of Integrals Variance Reduction TechniquesNonuniform Random Number Generation How to generate nonuniform random variables? Basic idea: Apply transformations to a stream of iid U[0,1] random variables Axel Gandy Monte Carlo Simulation 9
  9. 9. Random Number Generation Computation of Integrals Variance Reduction TechniquesInversion Method Let F be a cdf. Quantile function (essentially the inverse of the cdf): F −1 (x) = inf{x : F (x) ≥ u} If U is uniform then F −1 (U) ∼ F . Indeed, P(X ≤ x) = P(F −1 (U) ≤ x) = P(U ≤ F (x)) = F (x) Only works if F −1 (or a good approximation of it) is available. Axel Gandy Monte Carlo Simulation 10
  10. 10. Random Number Generation Computation of Integrals Variance Reduction TechniquesAcceptance-Rejection Method f(x) Cg(x) 0.30 0.20 0.10 0.00 0 5 10 15 20 x target density f Proposal density g (easy to generate from) such that for some C < ∞: f (x) ≤ Cg (x)∀x Algorithm: 1. Generate X from g . 2. With probability Cg(X )) return X - otherwise goto 1. f (X 1 C= probability of acceptance - want it to be as close to 1 as possible. Axel Gandy Monte Carlo Simulation 11
  11. 11. Random Number Generation Computation of Integrals Variance Reduction TechniquesFurther Algorithms Ratio-of-Uniforms Use of the characteristic function MCMC For many of those techniques and techniques to simulate specific distributions see e.g. Gentle (2003). Axel Gandy Monte Carlo Simulation 12
  12. 12. Random Number Generation Computation of Integrals Variance Reduction TechniquesEvaluation of an Integral Want to evaluate I := g (x)dx [0,1]d Importance for statistics: computation of expected values (posterior means), probabilities (p-values), variances, normalising constants, .... Often, d is large. In a random sample, often d =sample size. How to solve it? Symbolical Numerical Integration Quasi Monte Carlo Monte Carlo Integration Axel Gandy Monte Carlo Simulation 14
  13. 13. Random Number Generation Computation of Integrals Variance Reduction TechniquesNumerical Integration/Quadrature Main idea: approximate the function locally with simple function/polynomials Advantage: good convergence rate Not useful for high dimensions - curse of dimensionality Axel Gandy Monte Carlo Simulation 15
  14. 14. Random Number Generation Computation of Integrals Variance Reduction TechniquesMidpoint Formula Basic: 1 1 f (x)dx ≈ f 0 2 Composite: apply the rule in n subintervals 1.0 0.5 0.0 −0.5 −1.0 0.0 0.5 1.0 1.5 2.0 1 Error: O( n ). Axel Gandy Monte Carlo Simulation 16
  15. 15. Random Number Generation Computation of Integrals Variance Reduction TechniquesTrapezoidal Formula Basic: 1 1 f (x)dx ≈ (f (0) + f (1)) 0 2 Composite: 1.0 0.5 0.0 −0.5 −1.0 0.0 0.5 1.0 1.5 2.0 1 Error: O( n2 ). Axel Gandy Monte Carlo Simulation 17
  16. 16. Random Number Generation Computation of Integrals Variance Reduction TechniquesSimpson’s rule Approximate the integrand by a quadratic function 1 1 1 f (x)dx ≈ [f (0) + 4f ( ) + f (1)] 0 6 2 Composite Simpson: 1.0 0.5 0.0 −0.5 −1.0 0.0 0.5 1.0 1.5 2.0 1 Error: O( n4 ). Axel Gandy Monte Carlo Simulation 18
  17. 17. Random Number Generation Computation of Integrals Variance Reduction TechniquesAdvanced Numerical Integration Methods Newton Cotes formulas Adaptive methods Unbounded integration interval: transformations Axel Gandy Monte Carlo Simulation 19
  18. 18. Random Number Generation Computation of Integrals Variance Reduction TechniquesCurse of dimensionality - Numerical Integration in HigherDimensions I := g (x)dx [0,1]d Naive approach: write as iterated integral 1 1 I := ... g (x)dxn . . . dx1 0 0 use 1D scheme for each integral with, say g points . n = g d function evaluations needed for d = 100 (a moderate sample size) and g = 10 (which is not a lot): n > estimated number of atoms in the universe! 1 Suppose we use the trapezoidal rule, then the error = O( n2/d ) More advanced schemes are not doing much better! Axel Gandy Monte Carlo Simulation 20
  19. 19. Random Number Generation Computation of Integrals Variance Reduction TechniquesMonte Carlo Integration n 1 g (x)dx ≈ g (Xi ), [0,1]d n i=1 where X1 , X2 , · · · ∼ U([0, 1]d ) iid. SLLN: n 1 g (Xi ) → g (x)dx (n → ∞) n [0,1]d i=1 1 CLT: error is bounded by OP ( √n ). independent of d Can easily compute asymptotic confidence intervals. Axel Gandy Monte Carlo Simulation 21
  20. 20. Random Number Generation Computation of Integrals Variance Reduction TechniquesQuasi-Monte-Carlo Similar to MC, but instead of random Xi : Use deterministic xi that fill [0, 1]d evenly. so-called “low-discrepancy sequences”. R-package randtoolbox Axel Gandy Monte Carlo Simulation 22
  21. 21. Random Number Generation Computation of Integrals Variance Reduction TechniquesComparison between Quasi-Monte-Carlo and Monte Carlo- 2D 1000 Points in [0, 1]2 generated by a quasi-RNG and a Pseudo-RNG Quasi−Random−Number−Generator Standard R−Randon Number Generator 1.0 1.0 q q q q q q q qq q q q q q qq q q q qq qq q q qq q qq q q q q q q q q q qqq q q q q q q q q qq q q q q q q q q qq q qqq q q q qq q qq q q qqq q q q qq q q q q q q q q qq q q q q q q q qq q q q qq q q q qq q q q q q q q q qq q qq q q q q q q q q qq q qq q q q q q q q q qq q q q q q qq qq qq q q q q qq q q q q q q q q q q qq q q q q q qq q q q q q q qq qq q q q q q q q q qqq q q q q q q q q q q q q q qq q q qqq q q q q q q q q q q q q q q q q qq qq q q q qq q q q qq qq q qq q qq q q q q q q qq q q q qq q qq q qq q q qq q qq q q q q q qq q q qq q q q qq q q q q q q q qq q qqq q q qq q q qq qq q q q qq qq q qq q q q q q q q q qq q q q q q q qq q q q q q q q q q q q q q q q q q q q qqq q q q q q q qq q q q q q q qq q q q q qq q q qq q q qq q q 0.8 0.8 q qq q q qq q qq q q q q q q q q q q q q qq q qq q q q q q q q q q q q q q q q q q q q qq qq q q q q q qq q q q q q q q q q q q q q q q qq q q q q qq q q qq q q q qq q q qq q q q q q q qq q q q q q q q q q q q q q q q q q q qq q q q q q q qq qq qq q qqqqq q q qq qq q q q q qq qq q q q q q q q q q q q qq q q q qq q q qq q q q q q q q q q qq q q q qq q q qq q q qqqqq q q q q q qq q q q q q qq q q q q q q q q qq q qq q qq q q q q q q q q q q q q qq q q q q q q qq q q qq q qq q qq q q q q q q q qq q q qq q q q qq qq q q q qq q q q q q qq q q q q qq q q q qq q q q q qq qq q q qq q q q q qq q qq qq q q 0.6 0.6 q q q q q q qq q q q q q q q q q q qq q q q q q q qqq q q qq q q qq q q qq q q q q qq q q q qq q q qq q q q q qq q q q q qq q q qq q qq q q qqq q q q q qq qq qq q q q q q q q qq q q q q q q q q q q q q q qq q q q q q qqq q q q q q q q q q q qq q q qq q q qq qq q q q qq q q qq q q q q qq q q q qq q q q qq q qq q q q q q q q q q q q q q q q qq q q qq q q q q q q q qq q q qq q q qq q q qq q q q qq q q q qq q q q q q q q q q q qqq q qq qq q q qqqq q q q q q qq qq qq q q q qq q qq q q q qq qq q q q q q q q q q q q q q q q q qq q q q q q qq q q q q q q qq q q qq qq q q q q q q q q q q q q qqq qqqq q q q qq q q qq q q q q q q q q qq q q qq qq q q q qq q q q qq qq q q q q qq q qq q q q q q qqq q q q q q q q qq qq 0.4 0.4 q q q q q q q q qq q q q q qq q q q q q qq q q q q q q q q qq q q q q q qq q q q q qq q q q q q q q qq q q q qq q q q q q q qq q q q q q q q q qq qq q qq q q q q q q q qqq q q q q q q qq q q q q q q qq q q qq q q q qq qq q q q qq q q q q qq q q qq qq q qq q qq q q q qq q q qq q q q q q qq q q q qq q q q q q q qq qq q q q q qqqq q q q q q q q q q q qq q q q qq q qq q q q q q q q qq q q qq q q qq q q qq q q qq q q q qq qq q q q qq q q q q q q q q q q qq q q q q q q qq q qq q q q q qq q qq qq q q qq q q q q q q qq q q q q q q q qq q q qq q q q qq q q q q qq qq q qq q q q qq q q q qq q q qq q q q qq q qq q q q qq q q q q q q q qq q q qq q q q q q qq q q q qq q q q q q q qq q q q q q qq qq qq qq q q q q q q q qq q qq 0.2 0.2 q q q q q q q q q q qq q q qq qq q q qq qq q q q q q q q qq q q q q q q qq q q qq q qq qq q qq q q qq q q q q q q q q q q q q q qq q q qq qq q q q qq qq q q q q q q q q q q q qq q q q q q qq q q q q q q q qq q q q q q q qq qq q q q q q q q q q q q q q q qq q qq qqq q q qq q q q q q q q qq q qq q q q q q qq q q q q qq q q q q q qq q q q q qq q q q q q q qq q q q q qq q q q q q q q q qq q q q q qqq q q q q q q q q q q q q q q q qq q qqq q qq q q qq qq q q qq q q q q qq q q q q q q q qq q q qq q q qq qq q q q qq q q q q qq q q q qq qq q q q q q q qq qq q q q q q qq q q q q q qq qq q q q q q q q q qq q q qq q q q q q qq q q q q q q q q qq q q q q q qq q q qq qq q q q q q q q q qq q q q q q q q q q q q q q q q q q 0.0 0.0 q q q q q q q q q q q q q q q qq qq qq q q qq 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Axel Gandy Monte Carlo Simulation 23
  22. 22. Random Number Generation Computation of Integrals Variance Reduction TechniquesComparison between Quasi-Monte-Carlo and Monte Carlo (x1 + x2 )(x2 + x3 )2 (x3 + x4 )3 dx [0,1]4 Using Monte-Carlo and Quasi-MC 2.75 quasi−MC MC 2.70 2.65 2.60 2.55 0e+00 2e+05 4e+05 6e+05 8e+05 1e+06 n Axel Gandy Monte Carlo Simulation 24
  23. 23. Random Number Generation Computation of Integrals Variance Reduction TechniquesBounds on the Quasi-MC error Koksma-Hlawka inequality (Niederreiter, 1992, Theorem 2.11) 1 ∗ g (xi ) − g (x)dx ≤ Vd (g )Dn , n [0,1]d where Vd (g ) is the so-called Hardy and Krause variation of g and Dn is the discrepancy of the points x1 , . . . , xn in [0, 1]d given by ∗ Dn = sup |#{xi ∈ A : i = 1, . . . , n} − λ(A)| A∈A where λ is Lebesgue measure A is the set of all subrectangles of [0, 1]d of the d form i=1 [0, ai ]d Many sequences have been suggested, e.g. the Halton sequence (other sequences: Gandy Axel Faure, Sobol, ...) with:Carlo Simulation Monte 25
  24. 24. Random Number Generation Computation of Integrals Variance Reduction TechniquesComparison The consensus in the literature seems to be: use numerical integration for small d Quasi-MC useful for medium d use Monte Carlo integration for large d Axel Gandy Monte Carlo Simulation 26
  25. 25. Random Number Generation Computation of Integrals Variance Reduction TechniquesImportance Sampling Main idea: Change the density we are sampling from. Interested in E(φ(X )) = φ(x)f (x)dx For any density g , f (x) E(φ(X )) = φ(x) g (x)dx g (x) Thus an unbiased estimator of E(φ(X )) is n ˆ= 1 I φ(Xi ) f (Xi ) , n g (Xi ) i=1 where X1 , . . . , Xn ∼ g iid. How to choose g ? Suppose g ∝ φf then Var(ˆ) = 0. I However, the corresponding normalizing constant is E(φ(X )), the quantity we want to estimate! A lot of theoretical work is based on large deviation theory. Axel Gandy Monte Carlo Simulation 28
  26. 26. Random Number Generation Computation of Integrals Variance Reduction TechniquesImportance Sampling and Rare Events Importance sampling can greatly reduce the variance for estimating the probability of rare events, i.e. φ(x) = I(x ∈ A) and E(φ(X )) = P(X ∈ A) small. Axel Gandy Monte Carlo Simulation 29
  27. 27. Random Number Generation Computation of Integrals Variance Reduction TechniquesControl Variates Interested in I = E X Suppose we can also observe Y and know E Y . Consider T = X + a(Y − E(Y )) Then E T = I and Var T = Var X + 2a Cov(X , Y ) − a2 Var Y Minimized for a = − Cov(XY ) . Var ,Y usually, a not known → estimate For Monte Carlo sampling: generate iid sample (X1 , Y1 ), . . . , (Xn , Yn ) estimate Cov(X , Y ), Var Y based on this sample → ˆ a ˆ = 1 n [Xi + ˆ(Yi − E(Y ))] I a n i=1 ˆ can be computed via standard regression analysis. I Hence the term “regression-adjusted control variates” . Can be easily generalised to several control variates. Axel Gandy Monte Carlo Simulation 30
  28. 28. Random Number Generation Computation of Integrals Variance Reduction TechniquesFurther Techniques Antithetic Sampling Use X and −X Conditional Monte Carlo Evaluate parts explicitly Common Random Numbers For comparing two procedures - use the same sequence of random numbers. Stratification Divide sample space Ω into strata Ω1 , . . . , Ωs In each strata, generate Ri replicates conditional on Ωi and obtain an estimates ˆi I Combine using the law of total probability: ˆ = p1ˆ1 + · · · + ps ˆs I I I Need to know pi = P(Ωi ) for all i Axel Gandy Monte Carlo Simulation 31
  29. 29. Introduction Parallel RNG Practical use of parallel computing (R) Part II Parallel Computing Introduction Parallel RNG Practical use of parallel computing (R) Axel Gandy Parallel Computing 32
  30. 30. Introduction Parallel RNG Practical use of parallel computing (R)Moore’s Law (Source: Wikipedia, Creative Commons Attribution ShareAlike 3.0 License) Axel Gandy Parallel Computing 34
  31. 31. Introduction Parallel RNG Practical use of parallel computing (R)Growth of Data Storage Growth PC Harddrive Capacity q q q q q q q 1e+02 q q q q q q q q q q q q q q q q q q q q q q q q capacity [GB] q q q q q q q q q q q q q q q q q q q q q q q q q 1e+00 q q q q q q q q q q q q q q q q q q 1e−02 q q q q q q q q q q q q q q q q q q q q 1980 1985 1990 1995 2000 2005 2010 year Not only the computer speed but also the data size is increasing exponentially! The increase in the available storage is at least as fast as the increase in computing power. Axel Gandy Parallel Computing 35
  32. 32. Introduction Parallel RNG Practical use of parallel computing (R)Introduction Recently: Less increase in CPU clock speed → multi core CPUs are appearing (quad core available - 80 cores in labs) → software needs to be adapted to exploit this Traditional computing: Problem is broken into small steps that are executed sequentially Parallel computing: Steps are being executed in parallel Axel Gandy Parallel Computing 36
  33. 33. Introduction Parallel RNG Practical use of parallel computing (R)von Neumann Architecture CPU executes a stored program that specifies a sequence of read and write operations on the memory. Memory is used to store both program and data instructions Program instructions are coded data which tell the computer to do something Data is simply information to be used by the program A central processing unit (CPU) gets instructions and/or data from memory, decodes the instructions and then sequentially performs them. Axel Gandy Parallel Computing 37
  34. 34. Introduction Parallel RNG Practical use of parallel computing (R)Different Architectures Multicore computing Symmetric multiprocessing Distributed Computing Cluster computing Massive Parallel processor Grid Computing List of top 500 supercomputers at http://www.top500.org/ Axel Gandy Parallel Computing 38
  35. 35. Introduction Parallel RNG Practical use of parallel computing (R)Flynn’s taxonomy Single Instruction Multiple Instruction Single Data SISD MISD Multiple Data SIMD MIMD Examples: SIMD: GPUs Axel Gandy Parallel Computing 39
  36. 36. Introduction Parallel RNG Practical use of parallel computing (R)Memory Architectures of Parallel Computers Traditional System Memory CPU Shared Memory System Memory CPU CPU CPU Distributed Memory System Memory Memory Memory CPU CPU CPU Distributed Shared Memory System Memory Memory Memory CPU CPU CPU CPU CPU CPU CPU CPU CPU Axel Gandy Parallel Computing 40
  37. 37. Introduction Parallel RNG Practical use of parallel computing (R)Embarrassingly Parallel Computations Examples: Monte Carlo Integration Bootstrap Cross-Validation Axel Gandy Parallel Computing 41
  38. 38. Introduction Parallel RNG Practical use of parallel computing (R)Speedup Ideally: computational time reduced linearly in the number of CPUs Suppose only a fraction p of the total tasks can be parallelized. Supposing we have n parallel CPUs, the speedup is 1 (Amdahl’s Law) (1 − p) + p/n → no infinite speedup possible. Example: p = 90%, maximum speed up by a factor of 10. Axel Gandy Parallel Computing 42
  39. 39. Introduction Parallel RNG Practical use of parallel computing (R)Communication between processes Forking Threading OpenMP shared memory multiprocessing PVM (Parallel Virtual Machine) MPI (Message Passing Interface) How to divide tasks? e.g. Master/Slave concept Axel Gandy Parallel Computing 43
  40. 40. Introduction Parallel RNG Practical use of parallel computing (R)Parallel Random Number Generation Problems with RNG on parallel computers Cannot use identical streams Sharing a single stream: a lot of overhead. Starting from different seeds: danger of overlapping streams (in particular if seeding is not sophisticated or simulation is large) Need independent streams on each processor... Axel Gandy Parallel Computing 45
  41. 41. Introduction Parallel RNG Practical use of parallel computing (R)Parallel Random Number Generation - sketch of generalapproach 1 T 1 T 1 T ... s1 s2 s3 G G G f1 1 1 1 u1 u2 u3 f2 T T T s 2 s1 2 s2 2 s3 ... G G G f3 2 u1 2 u2 2 u3 3 T 3 T 3 T ... s1 s2 s3 G G G 3 u1 3 u2 3 u3 Axel Gandy Parallel Computing 46
  42. 42. Introduction Parallel RNG Practical use of parallel computing (R)Packages in R for Parallen random Number Generation rsprng Interface to the scalable parallel random number generators library (SPRNG) http://sprng.cs.fsu.edu/ rlecuyer Essentially starts with one random stream and partitions it into long substreams by jumping ahead. L’Ecuyer et al. (2002) Axel Gandy Parallel Computing 47
  43. 43. Introduction Parallel RNG Practical use of parallel computing (R)Profile Determine what part of the programme uses most time with a profiler Improve the important parts (usually the innermost loop) R has a built-in profiler (see Rprof, Rprof.summary, package profr) Axel Gandy Parallel Computing 49
  44. 44. Introduction Parallel RNG Practical use of parallel computing (R)Use Vectorization instead of Loops > a <- rnorm(1e+07) > system.time({ + x <- 0 + for (i in 1:length(a)) x <- x + a[i] + })[3] elapsed 21.17 > system.time(sum(a))[3] elapsed 0.07 Axel Gandy Parallel Computing 50
  45. 45. Introduction Parallel RNG Practical use of parallel computing (R)Just-In-Time Compilation - RA From the developer’s websie (http://www.milbo.users.sonic.net/ra/): “Ra is functionally identical to R but provides just-in-time compilation of loops and arithmetic expressions in loops. This usually makes arithmetic in Ra much faster. Ra will also typically run a little faster than standard R even when just-in-time compilation is not enabled.” Not just a package - central parts are reimplemented. Bill Venables (on R help archive): “if you really want to write R code as you might C code, then jit can help make it practical in terms of time. On the other hand, if you want to write R code using as much of the inbuilt operators as you have, then you can possibly still do things better.” Axel Gandy Parallel Computing 51
  46. 46. Introduction Parallel RNG Practical use of parallel computing (R)Use Compiled Code R is an interpreted language. Can include C, C++ and Fortran code. Can dramaticallly speed up computationally intensive parts (a factor of 100 is possible) No speedup if the computationally part is a vector/matrix operation. Downside: decreased portability Axel Gandy Parallel Computing 52
  47. 47. Introduction Parallel RNG Practical use of parallel computing (R)R-Package: snow Mostly for “embarassingly parallel” computations Extends the “apply”-style function to a cluster of machines: params <- 1:10000 cl <- makeCluster(8, "SOCK") res <- parSapply(cl, params, function(x) foo(x)) applies the function to each of the parameters using the cluster. will run 8 copies at once. Axel Gandy Parallel Computing 53
  48. 48. Introduction Parallel RNG Practical use of parallel computing (R)snow - Hello World > library(snow) > cl <- makeCluster(2, type = "SOCK") > str(clusterCall(cl, function() Sys.info()[c("nodename", "machine List of 2 $ : Named chr [1:2] "AG" "x86" ..- attr(*, "names")= chr [1:2] "nodename" "machine" $ : Named chr [1:2] "AG" "x86" ..- attr(*, "names")= chr [1:2] "nodename" "machine" > str(clusterApply(cl, 1:2, function(x) x + 3)) List of 2 $ : num 4 $ : num 5 > stopCluster(cl) Axel Gandy Parallel Computing 54
  49. 49. Introduction Parallel RNG Practical use of parallel computing (R)snow - set up random number generator without setting upt the RNG > cl <- makeCluster(2, type = "SOCK") > clusterApply(cl, 1:2, function(i) rnorm(5)) [[1]] [1] 0.1540537 -0.4584974 1.1320638 -1.4979826 1.1992120 [[2]] [1] 0.1540537 -0.4584974 1.1320638 -1.4979826 1.1992120 > stopCluster(cl) Now with proper setup of the RNG > cl <- makeCluster(2, type = "SOCK") > clusterSetupRNG(cl) [1] "RNGstream" > clusterApply(cl, 1:2, function(i) rnorm(5)) [[1]] [1] -1.14063404 -0.49815892 -0.76670013 -0.04821059 -1.09852152 [[2]] [1] 0.7049582 0.4821092 -1.2848088 0.7198440 0.7386390 > stopCluster(cl) Axel Gandy Parallel Computing 55
  50. 50. Introduction Parallel RNG Practical use of parallel computing (R)snow - Another Simple Example 5x4 processors - several servers > cl <- makeCluster(rep(c("localhost","euler","dirichlet","leibniz","riemann"), each=4),type="SOCK") may need to give password if not set up public/private key for ssh > system.time(sapply(1:1000,function(i) mean(rnorm(1e3))))[3] 0.156 > system.time(clusterApply(cl,1:1000,function(i) mean(rnorm(1e3))))[3] 0.161 > system.time(clusterApplyLB(cl,1:1000,function(i) mean(rnorm(1e3))))[3] 0.401 → too much overhead - parallelisation does not lead to gains > system.time(sapply(1:1000,function(i) mean(rnorm(1e5))))[3] 12.096 > system.time(clusterApply(cl,1:1000,function(i) mean(rnorm(1e5))))[3] 0.815 > system.time(clusterApplyLB(cl,1:1000,function(i) mean(rnorm(1e5))))[3] 0.648 > stopCluster(cl) → parallelisation leads to substantial gain in speed. Axel Gandy Parallel Computing 56
  51. 51. Introduction Parallel RNG Practical use of parallel computing (R) Extensions of snowsnowfall offers additional support for implicit sequential execution (e.g. for distributing packages using optional parallel support), additional calculation functions, extended error handling, and many functions for more comfortable programming.snowFT Extension of the snow package supporting fault tolerant and reproducible applications. It is written for the PVM communication layer. Axel Gandy Parallel Computing 57
  52. 52. Introduction Parallel RNG Practical use of parallel computing (R)Rmpi For more complicated parallel algorithms that are not embarassingly parallel. Tutorial under http://math.acadiau.ca/ACMMaC/Rmpi/ Hello world from this tutorial # Load the R MPI package if it is not already loaded. if (!is.loaded("mpi_initialize")) { library("Rmpi") } # Spawn as many slaves as possible mpi.spawn.Rslaves() # In case R exits unexpectedly, have it automatically clean up # resources taken up by Rmpi (slaves, memory, etc...) .Last <- function(){ if (is.loaded("mpi_initialize")){ if (mpi.comm.size(1) > 0){ print("Please use mpi.close.Rslaves() to close slaves.") mpi.close.Rslaves() } print("Please use mpi.quit() to quit R") .Call("mpi_finalize") } } # Tell all slaves to return a message identifying themselves mpi.remote.exec(paste("I am",mpi.comm.rank(),"of",mpi.comm.size())) # Tell all slaves to close down, and exit the program mpi.close.Rslaves() mpi.quit() (not able to install under win from CRAN - install from http://www.stats.uwo.ca/faculty/yu/Rmpi/) Axel Gandy Parallel Computing 58
  53. 53. Introduction Parallel RNG Practical use of parallel computing (R)Some other Packages nws Network of Workstations http://nws-r.sourceforge.net/ multicore Use of parallel computing on a single machine via fork (Unix, MacOS) - very fast and easy to use. GridR http: //cran.r-project.org/web/packages/GridR/ Wegener et al. (2009, Future Generation Computer Systems) papply on CRAN “ Similar to apply and lapply, applies a function to all items of a list, and returns a list with the results. Uses Rmpi to distribute the processing evenly across a cluster.”‘ multiR http://e-science.lancs.ac.uk/multiR/ rparallel http://www.rparallel.org/ Axel Gandy Parallel Computing 59
  54. 54. Introduction Parallel RNG Practical use of parallel computing (R)GPUs graphical processing units - in graphics cards very good at parallel processing need to taylor to specific GPU. Packages in R: gputools several basic routines. cudaBayesreg Bayesian multilevel modeling for fMRI. Axel Gandy Parallel Computing 60
  55. 55. Introduction Parallel RNG Practical use of parallel computing (R)Further Reading A tutorial on Parallel Computing: https://computing.llnl.gov/tutorials/parallel_comp/ High Performance Computing task view on CRAN http://cran.r-project.org/web/views/ HighPerformanceComputing.html An up-to-date talk on high performance comuting with R: http://dirk.eddelbuettel.com/papers/ useR2010hpcTutorial.pdf Axel Gandy Parallel Computing 61
  56. 56. References Part III Appendix Axel Gandy Appendix 62
  57. 57. ReferencesTopics in the coming lectures: Optimisation MCMC methods Bootstrap Particle Filtering Axel Gandy Appendix 63
  58. 58. ReferencesReferences I Gentle, J. (2003). Random Number Generation and Monte Carlo Methods. Springer. Knuth, D. (1981). The art of computer programming. Vol. 2: Seminumerical algorithms. Addison-Wesley. Knuth, G. (1998). The Art of Computer Programming, Seminumerical Algorithms, Vol.2 . L’Ecuyer, P. (1994). Uniform random number generation. Annals of Operations Research 53, 77–120. L’Ecuyer, P. (2004). Random number generation. In Handbook of Computational Statistics: Concepts and Methods, Springer. L’Ecuyer, P. (2006). Uniform random number generation. In Handbooks in Operations Research and Management Science, Elsevier. L’Ecuyer, P., Simard, R., Chen, E. J. & Kelton, W. D. (2002). An objected-oriented random-number package with many long streams and substreams. Operations Research 50, 1073–1075. The code in C, C++, Java, and FORTRAN is available. Axel Gandy Appendix 64
  59. 59. ReferencesReferences II Niederreiter, H. (1992). Random number generation and quasi-Monte Carlo methods. SIAM. Axel Gandy Appendix 65

×