122S:166Computing in StatisticsParallel computing in R using snowLect 27Nov.30, 2009Kate Cowles374 SH, [email protected] parallel jobs in R in the Linuxlab• use top and finger to make sure that yourtarget machines are not being used inten-sively by others• use the snow (“Simple Network of Worksta-tions”) package> library(snow)• master/slave design: master R process cre-ates a cluster of slave processes that carriesout computations and returns results to mas-ter3Main classes of functions in snow• makeCluster and stopCluster: start andstop a cluster of sl ave processes• clusterEvalQ: evalu a te a literal expressionon each slave and return results– useful f o r having each node run functionsusing arg uments i n its own memory• clusterCall: runs a specified function o nall slave nodes with identical arguments sentfrom master to all n odes• clusterApply and functions built on it: runsa specified function with arguments sent frommaster and split up among the nodes4Starting a cluster• makeCluster to start up a cluster of pro-cesses co mmunicating with master– node you started R in will be master# create an 8-process socket cluster> cl <- makeSOCKcluster(c("localhost","localhost","localhost","localhost","l-lnx211.divms.uiowa.edu","l-lnx211.divms.uiowa.edu","l-lnx211.divms.uiowa.edu","l-lnx211.divms.uiowa.edu"))# prompts for password for any machines I’m not already# logged into5clusterCall: to get all slave processes to exe-cute the same function with the same arguments# call Sys.info function on all processes> clusterCall(cl, function(cl) Sys.info() )> do.call("rbind", clusterCall(cl,function(cl) Sys.info()["nodename"]))nodename[1,] "l-lnx207.divms.uiowa.edu"[2,] "l-lnx207.divms.uiowa.edu"[3,] "l-lnx207.divms.uiowa.edu"[4,] "l-lnx207.divms.uiowa.edu"[5,] "l-lnx211.divms.uiowa.edu"[6,] "l-lnx211.divms.uiowa.edu"[7,] "l-lnx211.divms.uiowa.edu"[8,] "l-lnx211.divms.uiowa.edu"6Various parApply functions to get processes tocarry out same function on different arguments> mymat <- matrix(rnorm(1000000), nrow=1000)# calc summary statistics on each row of matrix using# master process only> system.time(applyout <- apply(mymat,1,summary))user system elapsed1.217 0.000 1.217> applyout <- t(applyout)> help(parRapply)7# calc summary statistics on each row of matrix using# 8 processes on 2 machines> system.time(parRapplyout1 <- parRapply(cl, mymat,summary))user system elapsed0.060 0.012 0.543# calc summary statistics on each row of matrix using# 4 processes on 1 machine> system.time(parRapplyout2 <- parRapply(cl[1:4], mymat,summary))user system elapsed0.089 0.006 0.4108> is.matrix(parRapplyout1)[1] FALSE> is.vector(parRapplyout1)[1] TRUE> parCapplyoutM <- matrix( parCapplyout, ncol=6, byrow=T)> parRapplyout1[1:12][1] -3.189000 -0.617100 0.022220 0.012210 0.652900 3.007000[8] -0.615500 -0.006234 0.005136 0.660500 3.563000> parRapplyout1M <- matrix(parRapplyout1, ncol=6, byrow=T )> parRapplyout1M[1:5, ]> applyout[1:5, ]Min. 1st Qu. Median Mean 3rd Qu. Max.[1,] -3.189 -0.6171 0.022220 0.012210 0.6529 3.007[2,] -2.887 -0.6155 -0.006234 0.005136 0.6605 3.563[3,] -2.833 -0.7610 -0.090580 -0.078740 0.5933 3.529[4,] -3.102 -0.6295 -0.051070 -0.016630 0.6048 3.774[5,] -3.535 -0.5880 0.099620 0.053010 0.7009 2.8159Parallel ra ndom number genera ti o n• If you don’t initialize a separate random num-ber stream for each node, it is possible to getidentical random number streams on all clus-ter nodes.• two parallel random number generators areavai lable in snow, based on two packages:rsprng and rlecuyer# seed independent uniform random number streams for snow cluster# L’Ecuyer’s generator> clusterSetupRNGstream(cl, seed=rep(12345,6))> do.call("rbind", clusterCall( cl, rnorm, 5 ))[,1] [,2] [,3] [,4] [,5][1,] 0.68567690 -1.40176821 0.4644450 0.06166349 0.3081519[2,] 1.59218721 0.98941360 -0.2260639 1.03891974 0.5725584[3,] -0.16213049 0.08560169 1.9705866 -1.10948803 1.7657923[4,] -0.57209565 1.79293554 0.9375416 -0.52962240 0.2293538[5,] 0.91600693 -0.23380300 -0.1648971 0.09438836 -1.1816592[6,] 0.05928533 0.01071665 -0.8184931 0.25428280 1.4123132[7,] 2.19466922 0.48531364 0.5833869 -1.18760882 -0.3608804[8,] -0.90408653 -0.32904130 0.5819193 0.19522605 1.139260210Getting individual processes to operate on dif-ferent data> stopCluster(cl) # I want a smaller cluster for# this example, so shut down 8-slave# cluster and make a smaller one> cl <- makeSOCKcluster(c("localhost","localhost","localhost","localhost"))> clusterSetupRNGstream(cl, seed=rep(12345,6))# make the working directory different for each slave process# set up a vector of arguments for clusterApply> dirnames <- c("dir1","dir2","dir3","dir4")> clusterApply(cl, dirnames, setwd )[[1]][1] "/mnt/nfs/netapp1/fs2/kcowles/166/parallel"[[2]][1] "/mnt/nfs/netapp1/fs2/kcowles/166/parallel"[[3]][1] "/mnt/nfs/netapp1/fs2/kcowles/166/parallel"[[4]][1] "/mnt/nfs/netapp1/fs2/kcowles/166/parallel"11> clusterCall( cl, getwd )[[1]][1] "/mnt/nfs/netapp1/fs2/kcowles/166/parallel/dir1"[[2]][1] "/mnt/nfs/netapp1/fs2/kcowles/166/parallel/dir2"[[3]][1] "/mnt/nfs/netapp1/fs2/kcowles/166/parallel/dir3"[[4]][1] "/mnt/nfs/netapp1/fs2/kcowles/166/parallel/dir4"# clusterEvalQ evaluates a literal expression on each node> clusterEvalQ( cl, load("y.Rdata") )[[1]][1] "y"[[2]][1] "y"[[3]][1] "y"[[4]][1] "y"12> clusterEvalQ( cl, print(y) ) # wouldn’t work with clusterCall[[1]][1] -0.3337495 -0.1518703 0.6569270 -0.7677969 -1.3420129[[2]][1] -0.5946356 1.0493249 2.1990838 -0.9631471 0.1113470 -1.1324455[[3]][1] 0.1571528 0.8139019 0.9223264 0.2692795[[4]][1] 0.86589506 1.27566060 -2.01340611 0.23612130 -0.56680754[7] -1.13304140 -0.64429002 0.02611008 -0.75593859> clusterEvalQ( cl, meany <- mean(y) )[[1]][1] -0.3877005[[2]][1] 0.1232192[[3]][1] 0.5406652[[4]][1] -0.276668713> clusterEvalQ( cl, ls() )[[1]][1] "meany" "y"[[2]][1] "meany" "y"[[3]][1] "meany" "y"[[4]][1] "meany" "y"> clusterEvalQ( cl, save(meany, file="meany.Rdata"))14clusterExport: Sending the same objec ts toall processes> x <- 2.5> wierdFunc <- function( x, y ){y * x }> clusterExport( cl, list("x", "wierdFunc") )> clusterEvalQ( cl, wierdFunc( x, y ) )[[1]][1] -0.8343737 -0.3796757 1.6423175 -1.9194923 -3.3550322[[2]][1] -1.4865890 2.6233122 5.4977095 -2.4078678 0.2783674 -2.8311138[[3]][1] 0.3928820 2.0347548 2.3058159 0.6731988[[4]][1] 2.1647377 3.1891515 -5.0335153 0.5903032 -1.4170189 -0.1424771[7] -2.8326035 -1.6107250 0.0652752 -1.889846515# important to shut
View Full Document
Unlocking...