I'm trying to adapt this R script for a speed test to work on a cluster.
When using the sfInit and makecluster functions with the type "SOCK", then the script runs on the cluster successfully, but without any speed improvement - unlike on my computer: when I change detectcores() to 1, the scripts runs substantially slower than with 4 cores.
I'm pretty sure I need to change the type to "MPI", though, in order to make the nodes communicate memory-wise with each other.
BUT: if I do so, the script then stops with following error code:
Loading required package: Rmpi
Error: package or namespace load failed for ‘Rmpi’:
.onLoad failed in loadNamespace() for 'Rmpi', details:
call: dyn.load(file, DLLpath = DLLpath, ...)
error: unable to load shared object '/cluster/sfw/R/3.5.1-gcc73-base/lib64/R/library/Rmpi/libs/Rmpi.so':
libmpi.so.20: cannot open shared object file: No such file or directory
Failed to load required library: Rmpi for parallel mode MPI
Fallback to sequential execution
snowfall 1.84-6.1 initialized: sequential execution, one CPU.
I thought "piece of cake, easy" and added the following lines:
install.packages('Rmpi', repos = "http://cran.us.r-project.org",
dependencies = TRUE, lib = '/personalpath') install.packages('doMPI',
repos = "http://cran.us.r-project.org", dependencies = TRUE, lib = '/personalpath') library(topicmodels, lib.loc = '/personalpath')
library(Rmpi, lib.loc = '/personalpath')
Which results in a successful installation but:
Error in library(Rmpi, lib.loc = "/personalpath") :
there is no package called ‘Rmpi’
1. How do I install these packages?
2. Do I really need to install them or is this a completely wrong approach?
Any help is highly appreciated! I know there are a couple of questions around here (see this, this, and this). But I'm not familiar with the calls in Linux and more importantly I do not have any rights on that cluster. So I need to come up with a solution in R...
So.. this is my code:
sfInit(parallel=TRUE, cpus=detectCores(), type="MPI")
cl <- makeCluster(detectCores(), type = "MPI")
registerDoSNOW(cl)
sfExport('dtm_stripped', 'control_LDA_Gibbs')
sfLibrary(topicmodels)
clusterEvalQ(cl, library(topicmodels))
clusterExport(cl, c("dtm_stripped", "control_LDA_Gibbs"))
BASE <- system.time(best.model.BASE <<- lapply(seq, function(d){LDA(dtm_stripped, control = control_LDA_Gibbs, method ='Gibbs', d)}))
PLYR_S <- system.time(best.model.PLYR_S <<- llply(seq, function(d){LDA(dtm_stripped, control = control_LDA_Gibbs, method ='Gibbs', d)}, .progress = "text"))
wrapper <- function (d) topicmodels:::LDA(dtm_stripped, control = control_LDA_Gibbs, method ='Gibbs', d)
PARLAP <- system.time(best.model.PARLAP <<- parLapply(cl, seq, wrapper))
DOPAR <- system.time(best.model.DOPAR <<- foreach(i = seq, .export = c("dtm_stripped", "control_LDA_Gibbs"), .packages = "topicmodels", .verbose = TRUE) %dopar% (LDA(dtm_stripped, control = control_LDA_Gibbs, method ='Gibbs', k=i)))
SFLAPP <- system.time(best.model.SFLAPP <<- sfLapply(seq, function(d){topicmodels:::LDA(dtm_stripped, control = control_LDA_Gibbs, method ='Gibbs', d)}))
SFCLU <- system.time(best.model.SFCLU <<- sfClusterApplyLB(seq, function(d){topicmodels:::LDA(dtm_stripped, control = control_LDA_Gibbs, method ='Gibbs', d)}))
PLYRP <- system.time(best.model.PLYRP <<- llply(seq, function(d){topicmodels:::LDA(dtm_stripped, control = control_LDA_Gibbs, method ='Gibbs', d)}, .parallel = TRUE))
results_speedtest <- rbind(BASE, PLYR_S, PARLAP, DOPAR, SFLAPP, SFCLU, PLYRP)
print(results_speedtest)
There are other ways to parallelize in R. Maybe this link will help, as the second page explains, what these cluster types such as socket, mpi and fork do: https://stat.ethz.ch/R-manual/R-devel/library/parallel/doc/parallel.pdf
Otherwise I can also recomment looking into the package
foreach, as syntax is a lot more like a regular for-loop. Note that some parallelizing packages not available for all operating systems.