Parallel Caret with doSNOW clusters on macOS and CentOS

Question

I'm following this tutorial Introduction to Machine Learning with R and caret (https://www.youtube.com/watch?v=z8PRU46I3NY) and get different machine behaviour when running R in parallel with doSNOW on macOS compared to centOS:

cl = makeCluster(4, type = 'SOCK')
registerDoSNOW(cl)

# build model
caret.cv = train(Survived ~ .,
                 data = titanic.train,
                 method = 'xgbTree',
                 tuneGrid = tune.grid,
                 trControl = train.control)
stopCluster(cl)

When running on macOS it creates 4 processes with each 1 thread so running 4@>99% (xgbTree in ~6min). On centOS it creates 4 processes each running 24 threads so in total 24@>99% (xgbTree not finishing >>30min). Even when only creating one or two clusters on centOS all threads are used and the server is completely busy.

UPDATE: When running non-caret code using doSNOW clusters everything works fine - running 1 thread per process, even on centOS.

Is there anything I'm missing? Should I expect different behaviour on these systems with identical scripts? Do I need to specify something for use on centOS?

I'm very new to caret & parallel R and so far I've read that there are only bigger differences between mac/linux and windows.

Please let me know if I can get you additional info. Thanks for your help and suggestions.

htop on centOS 60x+: R --slave --no-restore ==file=/usr/lib64/R/library/snow/RSOCKnode.R --args MASTER=localhost PORT=11326 OUT=/dev/null SNOWLIB=/usr/lib64/R/library

R Version 3.3.2: x86_64-redhat-linux-gnu ; x86_64-apple-darwin13.4.0 / centOS server: 2 sockets each 6 cores, each 2 threads / macOS MBP: 1/8/1

yes, when running on 4 clusters I get 4 processes, each with ~600%CPU usage. totalling 100% usage on all 24 threads — Daan11

Daan11 Daan11 · Accepted Answer · 2017-07-19T14:22:31

This solved my problem: Parallel processing with xgboost and caret

In contrast to my R/caret macOS installation it appears to be necessary to specify the number of threads (nthread = 1) for each xgboost process on the centOS installation:

caret.cv = train(yol ~ .,
             data = kmer.train,
             method = 'xgbTree',
             tuneGrid = tune.grid,
             trControl = train.control,
             nthread = 1)

While failing to do so will still result in 1 thread / process on macOS, xgboost will (as I understand) multithread and try to occupy all threads with every process.

Parallel Caret with doSNOW clusters on macOS and CentOS

1 Answers