A Future for BatchJobs

NOTE: The BatchJobs package is deprecated in favor of the batchtools package. Because of this, it is recommended to use the future.batchtools package instead of this package. This package will only be updated in order for it to be compliant with updates in R and the future package.

Introduction

The future package provides a generic API for using futures in R. A future is a simple yet powerful mechanism to evaluate an R expression and retrieve its value at some point in time. Futures can be resolved in many different ways depending on which strategy is used. There are various types of synchronous and asynchronous futures to choose from in the future package.

This package, future.BatchJobs, provides a type of futures that utilizes the BatchJobs package. This means that any type of backend that the BatchJobs package supports can be used as a future. More specifically, future.BatchJobs will allow you or users of your package to leverage the compute power of high-performance computing (HPC) clusters via a simple switch in settings - without having to change any code at all.

For instance, if BatchJobs is properly configures, the below two expressions for futures x and y will be processed on two different compute nodes:

> library("future.BatchJobs")
> plan(batchjobs_custom)
>
> x %<-% { Sys.sleep(5); 3.14 }
> y %<-% { Sys.sleep(5); 2.71 }
> x + y
[1] 5.85

This is obviously a toy example to illustrate what futures look like and how to work with them.

A more realistic example comes from the field of cancer research where very large data FASTQ files, which hold a large number of short DNA sequence reads, are produced. The first step toward a biological interpretation of these data is to align the reads in each sample (one FASTQ file) toward the human genome. In order to speed this up, we can have each file be processed by a separate compute node and each node we can use 24 parallel processes such that each process aligns a separate chromosome. Here is an outline of how this nested parallelism could be implemented using futures.

library("future")
library("listenv")
## The first level of futures should be submitted to the
## cluster using BatchJobs.  The second level of futures
## should be using multiprocessing, where the number of
## parallel processes is automatically decided based on
## what the cluster grants to each compute node.
plan(list(batchjobs_custom, multiprocess))

## Find all samples (one FASTQ file per sample)
fqs <- dir(pattern="[.]fastq$")

## The aligned results are stored in BAM files
bams <- listenv()

## For all samples (FASTQ files) ...
for (ss in seq_along(fqs)) {
  fq <- fqs[ss]

  ## ... use futures to align them ...
  bams[[ss]] %<-% {
    bams_ss <- listenv()
    ## ... and for each FASTQ file use a second layer
    ## of futures to align the individual chromosomes
    for (cc in 1:24) {
      bams_ss[[cc]] %<-% htseq::align(fq, chr=cc)
    }
    ## Resolve the "chromosome" futures and return as a list
    as.list(bams_ss)
  }
}
## Resolve the "sample" futures and return as a list
bams <- as.list(bams)

Note that a user who do not have access to a cluster could use the same script processing samples sequentially and chromosomes in parallel on a single machine using:

plan(list(sequential, multiprocess))

or samples in parallel and chromosomes sequentially using:

plan(list(multiprocess, sequential))

For an introduction as well as full details on how to use futures, please consult the package vignettes of the future package.

Choosing BatchJobs backend

The future.BatchJobs package implements a generic future wrapper for all BatchJobs backends. Below are the most common types of BatchJobs backends.

In addition to the above, there is also batchjobs_multicore (which on Windows and Solaris falls back to batchjobs_local), which runs BatchJobs tasks asynchronously in background R sessions (sic!) on the current machine. We advise to not use this and instead use multisession of the future package. For details, see help("batchjobs_multicore").

Examples

Below are two examples illustrating how to use batchjobs_custom and batchjobs_torque to configure the BatchJobs backend. For further details and examples on how to configure BatchJobs, see the BatchJobs configuration wiki page.

Example: A .BatchJobs.R file using local BatchJobs

The most general way of configuring BatchJobs is via a .BatchJobs.R file. This file should be located in the current directory or in the user's home directory. For example, as an alternative to batchjobs_local, we can manually configure local BatchJobs futures a .BatchJobs.R file that contains

cluster.functions <- makeClusterFunctionsLocal()

This will then be found and used when specifying

> plan(batchjobs_custom)

To specify this BatchJobs configuration file explicitly, one can use

> plan(batchjobs_custom, pathname="./.BatchJobs.R")

This follows the naming convention set up by the BatchJobs package.

Example: A .BatchJobs.*.tmpl template file for TORQUE / PBS

To configure BatchJobs for job schedulers we need to setup a template file that is used to generate the script used by the scheduler. This is what a template file for TORQUE / PBS may look like:

## Job name:
#PBS -N <%= job.name %>

## Merge standard error and output:
#PBS -j oe

## Resource parameters:
<% for (name in names(resources)) { %>
#PBS -l <%= name %>=<%= resources[[name]] %>
<% } %>

## Run R:
R CMD BATCH --no-save --no-restore "<%= rscript %>" /dev/stdout

If this template is saved to file .BatchJobs.torque.tmpl in the working directory or the user's home directory, then it will be automatically located and loaded when doing:

> plan(batchjobs_torque)

Resource parameters can be specified via argument resources which should be a named list and is passed as is to the template file. For example, to request that each job would get allotted 12 cores (one a single machine) and up to 5 GiB of memory, use:

> plan(batchjobs_torque, resources=list(nodes="1:ppn=12", vmem="5gb"))

To specify the resources argument at the same time as using nested future strategies, one can use tweak() to tweak the default arguments. For instance,

plan(list(
  tweak(batchjobs_torque, resources=list(nodes="1:ppn=12", vmem="5gb")),
  multiprocess
))

causes the first level of futures to be submitted via the TORQUE job scheduler requesting 12 cores and 5 GiB of memory per job. The second level of futures will be evaluated using multiprocessing using the 12 cores given to each job by the scheduler.

A similar filename format is used for the other types of job schedulers supported. For instance, for Slurm the template file should be named .BatchJobs.slurm.tmpl in order for

> plan(batchjobs_slurm)

to locate the file automatically. To specify this template file explicitly, use argument pathname, e.g.

> plan(batchjobs_slurm, pathname="./.BatchJobs.slurm.tmpl")

Note that it is still possible to use a .BatchJobs.R and load the template file using a standard BatchJobs approach for maximum control. For further details and examples on how to configure BatchJobs per se, see the BatchJobs configuration wiki page.

Demos

The future package provides a demo using futures for calculating a set of Mandelbrot planes. The demo does not assume anything about what type of futures are used. The user has full control of how futures are evaluated. For instance, to use local BatchJobs futures, run the demo as:

library("future.BatchJobs")
plan(batchjobs_local)
demo("mandelbrot", package="future", ask=FALSE)

Copyright Henrik Bengtsson, 2015-2018