Re: [R-pkg-devel] parallel computing slower than sequential computing
-Original Message- From: Ivan Krylov Sent: Tuesday, April 30, 2019 10:17 AM To: Wang, Zhu Cc: R-package-devel@r-project.org Subject: Re: [R-pkg-devel] parallel computing slower than sequential computing On Mon, 29 Apr 2019 23:44:42 + "Wang, Zhu" wrote: > sessionInfo() > R version 3.5.2 (2018-12-20) > Platform: x86_64-pc-linux-gnu (64-bit) Running under: Ubuntu 18.04.2 > LTS Which BLAS implementation do you use? One popular implementation, OpenBLAS, spawns multiple threads to do some operations faster; the threads can compete against each other for CPU resources if resulting number of processes * threads per process is more than what CPU can handle. >>BLAS: /usr/lib/x86_64-linux-gnu/blas/libblas.so.3.7.1 >>LAPACK: /usr/lib/x86_64-linux-gnu/lapack/liblapack.so.3.7.1 How many CPU cores does your system have? Does this include SMT (also known as hyper-threading on Intel processors)? While some problems benefit from processor pipeline being able to fetch from multiple threads at the same time, for others it's more of a bottleneck. >>8 CPU cores and 16 logical processors. The Linux system is on a Virtualbox. I >>realized this might be a factor. It may help to decrease the n.cores parameter. >> n.cores=3, 4, 5 would have similar user/elapsed time compared with >> sequential computing. Thanks Ivan. -- Best regards, Ivan __ R-package-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-package-devel
Re: [R-pkg-devel] parallel computing slower than sequential computing
On Mon, 29 Apr 2019 23:44:42 + "Wang, Zhu" wrote: > sessionInfo() > R version 3.5.2 (2018-12-20) > Platform: x86_64-pc-linux-gnu (64-bit) > Running under: Ubuntu 18.04.2 LTS Which BLAS implementation do you use? One popular implementation, OpenBLAS, spawns multiple threads to do some operations faster; the threads can compete against each other for CPU resources if resulting number of processes * threads per process is more than what CPU can handle. How many CPU cores does your system have? Does this include SMT (also known as hyper-threading on Intel processors)? While some problems benefit from processor pipeline being able to fetch from multiple threads at the same time, for others it's more of a bottleneck. It may help to decrease the n.cores parameter. -- Best regards, Ivan __ R-package-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-package-devel
[R-pkg-devel] parallel computing slower than sequential computing
Hello,
I have a package where parallel computing is slower than sequential computing
with cross-validation. For instance, please see the following reproducible
example with 10 cores and the relevant lines in the corresponding R function.
The R package is also attached. I thought parallel computing should improve the
speed. However, it is not obvious to me where it went wrong.
Many thanks in advance.
Zhu Wang
library("mpath")
library("pscl")
data("bioChemists", package = "pscl")
>system.time(cv.zipath(art ~ . | ., data = bioChemists, family = "negbin",
>nlambda=100, parallel=FALSE))
user system elapsed
77.430 0.031 79.547
> system.time(cv.zipath(art ~ . | ., data = bioChemists, family = "negbin",
> nlambda=100, parallel=TRUE, n.cores=10))
user system elapsed
95.694 0.517 106.072
R> sessionInfo()
R version 3.5.2 (2018-12-20)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Ubuntu 18.04.2 LTS
The relevant lines in cv.zipath regarding parallel argument are below:
if(parallel){
registerDoParallel(cores=n.cores)
i <- 1 ###needed to pass R CMD check with parallel code below
residmat <- foreach(i=seq(K), .combine=cbind, .packages='mpath') %dopar% {
omit <- all.folds[[i]]
fitcv <- do.call("zipath", list(formula, data[-omit,], weights[-omit],
lambda.count=lambda.count, lambda.zero=lambda.zero, nlambda=nlambda, ...))
logLik(fitcv, newdata=data[omit,, drop=FALSE], Y[omit],
weights=weights[omit])
}
stopImplicitCluster()
}
else{
residmat <- matrix(NA, nlambda, K)
for(i in seq(K)) {
if(trace)
cat("\n CV Fold", i, "\n\n")
omit <- all.folds[[i]]
fitcv <- do.call("zipath", list(formula, data[-omit,], weights[-omit],
lambda.count=lambda.count, lambda.zero=lambda.zero, nlambda=nlambda, ...))
residmat[, i] <- logLik(fitcv, newdata=data[omit,, drop=FALSE],
Y[omit], weights=weights[omit])
}
}
mpath_0.3-13.tar.gz
Description: mpath_0.3-13.tar.gz
__
R-package-devel@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-package-devel
