Difference between revisions of "R parallel"

From 太極
Jump to navigation Jump to search
 
(One intermediate revision by the same user not shown)
Line 169: Line 169:
 
</syntaxhighlight>
 
</syntaxhighlight>
  
== mclapply() (Windows OS only supports 1 core) vs parLapply()/parSapply() (all OSs) ==
+
== mclapply() (Windows OS only supports 1 core, forking) vs parLapply()/parSapply() (all OSs, clustering) ==
 
* Be providing the number of cores in mclapply() using '''mc.cores''' argument (2 is used by default)
 
* Be providing the number of cores in mclapply() using '''mc.cores''' argument (2 is used by default)
 
* set.seed() can not fix seed in mclapply(); we need to use set.seed(, "L'Ecuyer-CMRG"). But be careful on the side-effect of using '''L'Ecuyer-CMRG''' seed.  
 
* set.seed() can not fix seed in mclapply(); we need to use set.seed(, "L'Ecuyer-CMRG"). But be careful on the side-effect of using '''L'Ecuyer-CMRG''' seed.  
Line 242: Line 242:
 
** ''mclapply''() uses 10.5GB while ''parLapply''() uses around 6.5GB and ''foreach''() (either doParallel or doSNOW backend) uses about 7GB (the numbers are obtained by monitoring the output from ''jobload'' command). NB: the memory used from the output of ''jobhist'' command in Biowulf is not reliable
 
** ''mclapply''() uses 10.5GB while ''parLapply''() uses around 6.5GB and ''foreach''() (either doParallel or doSNOW backend) uses about 7GB (the numbers are obtained by monitoring the output from ''jobload'' command). NB: the memory used from the output of ''jobhist'' command in Biowulf is not reliable
 
** They are taking about the same amount of running time (19 minutes)
 
** They are taking about the same amount of running time (19 minutes)
** mclapply() does not use ''clusterExport''() nor ''clusterEvalQ''() while others need to
+
** mclapply() uses fork. So it does not need ''clusterExport''() nor ''clusterEvalQ''() while other methods like parLapply() need.
 
** '''Bottomline: Do not use mclapply()''' if we want to save memory
 
** '''Bottomline: Do not use mclapply()''' if we want to save memory
 
** The code is available in my private bitbucket repository
 
** The code is available in my private bitbucket repository

Latest revision as of 09:15, 18 May 2020

Introduction

Security warning from Windows/Mac

It seems it is safe to choose 'Cancel' when Windows Firewall tried to block R program when we use makeCluster() to create a socket cluster.

library(parallel)
cl <- makeCluster(2)
clusterApply(cl, 1:2, get("+"), 3)
stopCluster(cl)

WindowsSecurityAlert.png RegisterDoParallel mac.png

If we like to see current firewall settings, just click Windows Start button, search 'Firewall' and choose 'Windows Firewall with Advanced Security'. In the 'Inbound Rules', we can see what programs (like, R for Windows GUI front-end, or Rserve) are among the rules. These rules are called 'private' in the 'Profile' column. Note that each of them may appear twice because one is 'TCP' protocol and the other one has a 'UDP' protocol.

Detect number of cores and memory consideration

parallel::detectCores()  # not good
future::availableCores() # good

Don't use the default option getOption("mc.cores", 2L) (PS it only returns 2.) in mclapply() unless you are a developer for a package.

However, it is a different story when we run the R code in HPC cluster. Read the discussion Whether to use the detectCores function in R to specify the number of cores for parallel processing?

On NIH's biowulf, even I specify an interactive session with 4 cores, the parallel::detectCores() function returns 56. This number is the same as the output from the bash command grep processor /proc/cpuinfo or (better) lscpu. The free -hm also returns a full 125GB size instead of my requested size (4GB by default). The future::availableCores() fixes the problem.

See Biowulf's R webpage for a detailed instruction and my note Running R scripts for selecting the number of cores on biowulf/non-biowulf (different OS).

Identify OS

See Distinguish Windows, Linux, Mac

Memory

  • Alternative to R's `memory.size()` in linux? pryr::mem_used() will return the memory currently used by R.
  • system('grep MemTotal /proc/meminfo', intern = TRUE) will show the total physical RAM on Linux OS (mac does not work).
  • Monitor memory usage with R parallel. profvis package from RStudio.
    • It works on RStudio or a regular R session
    • The output is an HTML file
    • We can change the interval period. Its default is 0.01 seconds or 100 ms. profvis(expr = NULL, interval = 0.01, ...)
    • When I test profvis() on a function which calls clusterExport() and parSapply(), I can see the time spent on each of them. My example of running coxph on a set of genes.
    • Question: why the sum of time from the HTML output not close at all to system.time() when parLapply() is used?
    • Advanced R > Measuring performance
    • Efficient R programming > Code profiling.
  • Rprof() and summaryRprof()
Rprof(tf <- "rprof.txt", memory.profiling = TRUE)
# the code you want to profile must be in between
Rprof (NULL) ; print(summaryRprof(tf))

Socket or Fork

Parallel R: Socket or Fork

  • For the fork, each parallel thread is a complete duplication of the master process with the shared environment, including objects or variables defined prior to the kickoff of parallel threads. Therefore, it runs fast. However, the major limitation is that the fork doesn’t work on the Windows system.
  • The socket works on all operating systems. Each thread runs separately without sharing objects or variables, which can only be passed from the master process explicitly. As a result, it runs slower due to the communication overhead.
install.packages(c("rbenchmark", "nycflights13"))
data(flights, package = 'nycflights13')
dim(flights) # [1] 336776     19
ex <- expression(carrier == "UA" & origin == "EWR" & day == 1 & is.na(arr_time))

parFilter <- function(df, ex, type) {
  cn <- parallel::detectCores() - 1
  cl <- parallel::makeCluster(cn, type = type)
  ### DIVIDE THE DATAFRAME BASED ON # OF CORES
  sp <- parallel::parLapply(cl, parallel::clusterSplit(cl, seq(nrow(df))),
                            function(c_) df[c_,])
  ### PASS THE OBJECT FROM MASTER PROCESS TO EACH NODE
  parallel::clusterExport(cl, "ex")
  ### EXTRACT ROW INDEX ON EACH NODE
  id <- Reduce(c, parallel::parLapply(cl, sp,
                                      function(s_) with(s_, eval(ex))))
  parallel::stopCluster(cl)
  return(df[which(id),])
}

rbenchmark::benchmark(replications = 10, order = "elapsed", relative = "elapsed",
                      columns = c("test", "replications", "elapsed", "relative"),
                      "  FORK" = parFilter(flights, ex, "FORK"),
                      "SOCKET" = parFilter(flights, ex, "PSOCK")
)  # Phenom(tm) II X6 1055T
    test replications elapsed relative
1   FORK           10  42.154    1.000
2 SOCKET           10  68.901    1.635

makeCluster() for socket connection in R 4.0

The function hangs on R 4.0.0 + RStudio. See

How can I print when using %dopar%

https://stackoverflow.com/a/15078540

Create and register your cluster with something like:

library(doParallel)
cl<-makeCluster(4, outfile = "debugParallel.txt")
registerDoParallel(cl)
foreach(x=list(1, 2, "a"))  %dopar%  
{
  print(x)
}
stopCluster(cl)

And the debugParallel.txt will look like

starting worker pid=17718 on localhost:11704 at 15:10:29.665
starting worker pid=17733 on localhost:11704 at 15:10:29.850
starting worker pid=17747 on localhost:11704 at 15:10:30.025
starting worker pid=17761 on localhost:11704 at 15:10:30.183
[1] 1
[1] 2
[1] "a"

Setting outfile to the empty string ("") will prevent doParallel (or snow whatever) from redirecting the output, often resulting in the output from your print messages showing up on the terminal of the master process. Your foreach loop doesn't need to change at all.

parallel package (including parLapply, parSapply)

Parallel package was included in R 2.14.0. It is derived from the snow and multicore packages and provides many of the same functions as those packages.

The parallel package provides several *apply functions for R users to quickly modify their code using parallel computing.

  • makeCluster(makePSOCKcluster, makeForkCluster), stopCluster. Other cluster types are passed to package snow.
  • clusterCall, clusterEvalQ: source R files and/or load libraries
  • clusterSplit
  • clusterApply, clusterApplyLB (vs the foreach package)
  • clusterExport: export variables. We may need to apply the option envir=environment() if the objects to be exported are local variables.
  • clusterMap
  • parallel::mclapply() and parallel::parLapply()
  • parLapply, parSapply, parApply, parRapply, parCapply. Note that
    • parSapply() can be used to as a parallel version of the replicate() function. See this example or this example.
    • An iteration parameter needs to be added to the first parameter of the main function.
  • parLapplyLB, parSapplyLB (load balance version)
  • clusterSetRNGStream, nextRNGStream, nextRNGSubStream

Examples (See ?clusterApply)

library(parallel)
cl <- makeCluster(2, type = "SOCK")
clusterApply(cl, 1:2, function(x) x*3)    # OR clusterApply(cl, 1:2, get("*"), 3)
# [[1]]
# [1] 3
#
# [[2]]
# [1] 6
parSapply(cl, 1:20, get("+"), 3)
#  [1]  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
stopCluster(cl)

An example of using clusterCall() or clusterEvalQ()

library(parallel)

cl <- makeCluster(4)
clusterCall(cl, function() { 
  source("test.R") 
})
# clusterEvalQ(cl, {
#    source("test.R")
# })

## do some parallel work
stopCluster(cl)

mclapply() (Windows OS only supports 1 core, forking) vs parLapply()/parSapply() (all OSs, clustering)

  • Be providing the number of cores in mclapply() using mc.cores argument (2 is used by default)
  • set.seed() can not fix seed in mclapply(); we need to use set.seed(, "L'Ecuyer-CMRG"). But be careful on the side-effect of using L'Ecuyer-CMRG seed.
  • Chapter 4 of Parallel R by Stephen Weston, Q. Ethan McCallum
  • Setting a seed in R, when using parallel simulation
  • R doesn't reset the seed when “L'Ecuyer-CMRG” RNG is used?
    library(parallel)
    system.time(mclapply(1:1e4L, function(x) rnorm(x)))
    system.time(mclapply(1:1e4L, function(x) rnorm(x), mc.cores = 4))
    
    set.seed(1234)
    mclapply(1:3, function(x) rnorm(x))
    set.seed(1234)
    mclapply(1:3, function(x) rnorm(x)) # cannot reproduce the result
    
    mclapply(1:3, function(x) {set.seed(x); rnorm(x)}) # Can reproduce the result
    
    set.seed(123, "L'Ecuyer")
    mclapply(1:3, function(x) rnorm(x))
    mclapply(1:3, function(x) rnorm(x)) # results are not changed once we have run set.seed( , "L'Ecuyer")
    
    set.seed(1234)                      # use set.seed() in order to get a new reproducible result
    mclapply(1:3, function(x) rnorm(x)) 
    mclapply(1:3, function(x) rnorm(x)) # results are not changed
    
    # an example from ?mclapply
    library(boot)
    cd4.rg <- function(data, mle) MASS::mvrnorm(nrow(data), mle$m, mle$v)
    cd4.mle <- list(m = colMeans(cd4), v = var(cd4))
    mc <- getOption("mc.cores", 2)
    run1 <- function(...) boot(cd4, corr, R = 500, sim = "parametric",
                               ran.gen = cd4.rg, mle = cd4.mle)
    ## To make this reproducible:
    set.seed(123, "L'Ecuyer")
    res <- mclapply(seq_len(mc), run1)
    cd4.boot <- do.call(c, res)
    boot.ci(cd4.boot,  type = c("norm", "basic", "perc"),
            conf = 0.9, h = atanh, hinv = tanh)
    
  • An example of applying to the permutation test (LS test for GSEA)
ngenes <- 10000
nresamp <- 1e5
genesetSize <- 20
pinter <- runif(ngenes)
lsobs <- -mean(log(pinter[genesetSize]))
system.time( lsperm <- mclapply(1:nresamp, function(i) -mean(log(sample(pinter, genesetSize)))) )
# system.time( lsperm <- mclapply(1:nresamp, function(i) -mean(log(sample(pinter, genesetSize))), mc.cores = 4) )
lsperm <- unlist(lsperm)
mean(lsperm > lsobs)
  • tempdir() is per session only. So the following will give the same directory.
parallel::mclapply(1:3, function(x) tempdir())

Note

  1. Windows OS can not use mclapply(). The mclapply() implementation relies on forking and Windows does not support forking. mclapply from the parallel package is implemented as a serial function on Windows systems. The parallelsugar package was created based on the above idea.
  2. Another choice for Windows OS is to use parLapply() function in parallel package.
  3. Understanding the differences between mclapply and parLapply in R You don't have to worry about reproducing your environment on each of the cluster workers if mclapply() is used.
    ncores <- as.integer( Sys.getenv('NUMBER_OF_PROCESSORS') )
    cl <- makeCluster(getOption("cl.cores", ncores))
    LLID2GOIDs2 <- parLapply(cl, rLLID, function(x) {
                                        library(org.Hs.eg.db); get("org.Hs.egGO")[[x]]} 
                            )
    stopCluster(cl)
    
    It does work. Cut the computing time from 100 sec to 29 sec on 4 cores.

mclapply() vs parLapply() vs foreach()

clusterApply() vs parLapply() and load-balancing versions

Look at the documentation clusterApply, it seems there is no difference in these two.

Note that there are load-balancing versions of these two: clusterApplyLB() and parLapplyLB().

parallel vs doParallel package

doParallel is a foreach parallel adaptor for the parallel package.

doParallel provides a parallel backend for the %dopar% function using the parallel package.

doParallel depends on foreach, iterators, parallel and utils packages.

parallelsugar package

If we load parallelsugar, the default implementation of parallel::mclapply, which used fork based clusters, will be overwritten by parallelsugar::mclapply, which is implemented with socket clusters.

library(parallel) 

system.time( mclapply(1:4, function(xx){ Sys.sleep(10) }) )
##    user  system elapsed 
##    0.00    0.00   40.06 

library(parallelsugar)
## 
## Attaching package: ‘parallelsugar’
## 
## The following object is masked from ‘package:parallel’:
## 
##     mclapply

system.time( mclapply(1:4, function(xx){ Sys.sleep(10) }) )
##    user  system elapsed 
##    0.04    0.08   12.98

pbmcapply

pbmcapply: Tracking the Progress of Mc*pply with Progress Bar.

snow package

http://cran.r-project.org/web/packages/snow/index.html

Supported cluster types are "SOCK", "PVM", "MPI", and "NWS".

multicore package

This package is removed from CRAN.

Consider using package ‘parallel’ instead.

foreach package

Changes in the foreach package version 1.4.8, Feb 2020.

Simple example

This package depends on one of the following

  • doParallel - Foreach parallel adaptor for the parallel package
  • doSNOW - Foreach parallel adaptor for the snow package
  • doMC - Foreach parallel adaptor for the multicore package. Used in glmnet vignette.
  • doMPI - Foreach parallel adaptor for the Rmpi package
  • doRedis - Foreach parallel adapter for the rredis package

as a backend.

library(foreach)
library(doParallel)

m <- matrix(rnorm(9), 3, 3)

cl <- makeCluster(2, type = "SOCK")
registerDoParallel(cl) # register the parallel backend with the foreach package
foreach(i=1:nrow(m), .combine=rbind) %dopar%
  (m[i,] / mean(m[i,]))

stopCluster(cl)

See also this post Updates to the foreach package and its friends on Oct 2015.

combine list of lists

comb <- function(...) {
  mapply('cbind', ..., SIMPLIFY=FALSE)
}

library(foreach)
library(doParallel)

cl <- makeCluster(2)
registerDoParallel(cl) # register the parallel backend with the foreach package

m <- rbind(rep(1,3), rep(2,3))

# nrow(m) can represents number of permutations (2 in this toy example)
tmp <- foreach(i=1:nrow(m)) %dopar% {
  a <- m[i, ]
  b <- a * 10
  list(a, b)
}; tmp
# [[1]]
# [[1]][[1]]
# [1] 1 1 1
#
# [[1]][[2]]
# [1] 10 10 10
#
#
# [[2]]
# [[2]][[1]]
# [1] 2 2 2
#
# [[2]][[2]]
# [1] 20 20 20

foreach(i=1:nrow(m), .combine = "comb") %dopar% {
  a <- m[i,]
  b <- a * 10
  list(a, b)
}
# [[1]]
#      [,1] [,2]
# [1,]    1    2
# [2,]    1    2
# [3,]    1    2
#
# [[2]]
#      [,1] [,2]
# [1,]   10   20
# [2,]   10   20
# [3,]   10   20
stopCluster(cl)

Replacing double loops

library(foreach)
library(doParallel)

nc <- 4
nr <- 2

cores=detectCores()
cl <- makeCluster(cores[1]-1)
registerDoParallel(cl)
# set.seed(1234) # not work
# set.seed(1234, "L'Ecuyer-CMRG") # not work either
# library("doRNG")
# registerDoRNG(seed = 1985)    # not work with nested foreach
# Error in list(e1 = list(args = (1:nr)(), argnames = "i", evalenv = <environment>,  :
#  nested/conditional foreach loops are not supported yet.
m <- foreach (i = 1:nr, .combine='rbind') %:% # nesting operator
  foreach (j = 1:nc) %dopar% {
    rnorm(1, i*5, j) # code to parallelise
}
m
stopCluster(cl)

Note that since the random seed (see the next session) does not work on nested loop, it is better to convert nested loop (two indices) to a single loop (one index).

set seed and doRNG package

Export libraries, variables, functions

clusterEvalQ(cl, {
  library(biospear)
  library(glmnet)
  library(survival)
})
clusterExport(cl, list("var1", "foo2"))

If we use makeCluster(, "FORK"), then we don't need to do these.

Summary the result

foreach returns the result in a list. For example, if each component is a matrix we can use

  • Reduce("+", res)/length(res) # Reduce("+", res, na.rm = TRUE) not working
  • apply(simplify2array(res), 1:2, mean, na.rm = TRUE)

to get the average of matrices over the list.

Read a list files

Reading List Faster With parallel, doParallel, and pbapply

Don't use too many cores: Error in unserialize(socklist[[n]])

error reading from connection. See https://stackoverflow.com/a/24352032

Why is foreach() sometimes slower than for?

See https://stackoverflow.com/a/10414280 or https://stackoverflow.com/q/16963808

  • foreach is only advisable if you have relatively few rounds through very time consuming functions.
  • foreach() is best when the number of jobs does not hugely exceed the number of processors you will be using.
library(foreach)
library(parallel)
nCores <- future::availableCores()
cl <- makeCluster(nCores[[1]])
registerDoParallel(cl)

B<-10000

myFunc<-function() for(i in 1:B) sqrt(i)

myFunc2<-function() foreach(i = 1:B)  %do% sqrt(i)

myParFunc<-function() foreach(i = 1:B) %dopar% sqrt(i)

system.time(myFunc())
#   user  system elapsed 
#  0.001   0.002   0.004 
system.time(myFunc2())
#   user  system elapsed 
#  2.473   0.005   2.477 
system.time(myParFunc())
#   user  system elapsed 
#  3.560   0.242   3.829 

stopCluster(cl)

Another example: reduce the probe sets corresponding to the same gene symbol by taking the probe set with the largest gene expression. It is not just the speed problem; it runs out of memory on a 44928x393 matrix with a PC has 8GB RAM. On a smaller data with 5000 genes and 100 arrays it took 0.071 seconds using split() + sapply() but it took 39 seconds using the foreach method.

require(magrittr)
x <- matrix(rnorm(5000*100), nr=5000) # gene expression
y <- sample(3000, 5000, replace = TRUE) # gene symbol
rownames(x) <- names(y) <- 1:5000 # probe sets name

spt <- apply(x,1,mean,na.rm=TRUE) %>% split(., y)
ind <- sapply(spt, function(a) names(which.max(a)))
x.reduced <- x[ind, ]; rownames(x.reduced) <- names(spt)

Progress bar (doSNOW + foreach)

doParallel doSNOW
require(doParallel)
cores <- parallel::detectCores()
cl <- makeCluster(cores)
registerDoParallel(cl)
opts <- list(progress=progress)
result <- foreach(i=1:N) %dopar% {}
stopCluster(cl)
require(doSNOW)
require(Kendall)
cores <- parallel::detectCores() 
cl <- makeSOCKcluster(cores) 
registerDoSNOW(cl) 
mydata <- matrix(rnorm(8000*500), ncol=500) 
pb <- txtProgressBar(min=1, max=8000, style=3)
progress <- function(n) setTxtProgressBar(pb, n)
opts <- list(progress=progress)
result <- foreach(i=1:8000, .packages="Kendall", 
          .options.snow=opts, .combine='rbind') %dopar% {
  abc <- MannKendall(mydata[i,])
  data.frame(tau=abc$tau, sl=abc$sl)
}
close(pb)
stopCluster(cl)

snowfall package

http://www.imbi.uni-freiburg.de/parallel/docs/Reisensburg2009_TutParallelComputing_Knaus_Porzelius.pdf

Rmpi package

Some examples/tutorials

OpenMP

BiocParallel

RcppParallel

future & future.apply & doFuture packages

Apache Spark

Microsoft R Server

GPU

Threads

Benchmark

R functions to run timing

# Method 1
system.time( invisible(rnorm(10000)))

# Method 2
btime <- Sys.time()
invisible(rnorm(10000))
Sys.time() - btime # difftime(Sys.time(), btime)