BiocParallelParam

BiocParallelParam instances configure different parallel evaluation environments. Creating or register() ing a ‘Param’ allows the same code to be used in different parallel environments without a code re-write. Params listed are supported on all of Unix, Mac and Windows except MulticoreParam which is Unix and Mac only.

• SerialParam:

  Supported on all platforms.

  Evaluate BiocParallel-enabled code with parallel evaluation disabled. This approach is useful when writing new scripts and trying to debug code.

• MulticoreParam:

  Supported on Unix and Mac. On Windows, MulticoreParam dispatches to SerialParam.

  Evaluate BiocParallel-enabled code using multiple cores on a single computer. When available, this is the most efficient and least troublesome way to parallelize code. Windows does not support multi-core evaluation (the MulticoreParam object can be used, but evaluation is serial). On other operating systems, the default number of workers equals the value of the global option mc.cores (e.g.,getOption("mc.cores") ) or, if that is not set, the number of cores returned by arallel::detectCores() - 2 ; when number of cores cannot be determined, the default is 1.

  MulticoreParam uses ‘forked’ processes with ‘copy-on-change’ semantics – memory is only copied when it is changed. This makes it very efficient to invoke compared to other back-ends.

  There are several important caveats to using MulticoreParam. Forked processes are not available on Windows. Some environments, e.g., RStudio, do not work well with forked processes, assuming that code evaluation is single-threaded. Some external resources, e.g., access to files or data bases, maintain state in a way that assumes the resource is accessed only by a single thread. A subtle cost is that R’s garbage collector runs periodically, and ‘marks’ memory as in use. This effectively triggers a copy of the marked memory. R’s generational garbage collector is triggered at difficult-to-predict times; the effect in a long-running forked process is that the memory is eventually copied. See this post for additional details.

  MulticoreParam is based on facilities originally implemented in the multicore package and subsequently the parallel package in base.

• SnowParam:

  Supported on all platforms.

  Evaluate BiocParallel-enabled code across several distinct instances, on one or several computers. This is a straightforward approach for executing parallel code on one or several computers, and is based on facilities originally implemented in the snow package. Different types of snow ‘back-ends’ are supported, including socket and MPI clusters.

• BatchtoolsParam:

  Applicable to clusters with formal schedulers.

  Evaluate BiocParallel-enabled code by submitting to a cluster scheduler like SGE.

• DoparParam:

  Supported on all platforms.

  Register a parallel back-end supported by the foreach package for use with BiocParallel.

The simplest illustration of creating BiocParallelParam is

serialParam <- SerialParam()
serialParam

## class: SerialParam
##   bpisup: FALSE; bpnworkers: 1; bptasks: 0; bpjobname: BPJOB
##   bplog: FALSE; bpthreshold: INFO; bpstopOnError: TRUE
##   bpRNGseed: ; bptimeout: NA; bpprogressbar: FALSE
##   bpexportglobals: FALSE; bpexportvariables: FALSE; bpforceGC: FALSE
##   bpfallback: FALSE
##   bplogdir: NA
##   bpresultdir: NA

Most parameters have additional arguments influencing behavior, e.g., specifying the number of ‘cores’ to use when creating a MulticoreParam instance

multicoreParam <- MulticoreParam(workers = 8)
multicoreParam

## class: MulticoreParam
##   bpisup: FALSE; bpnworkers: 8; bptasks: 0; bpjobname: BPJOB
##   bplog: FALSE; bpthreshold: INFO; bpstopOnError: TRUE
##   bpRNGseed: ; bptimeout: NA; bpprogressbar: FALSE
##   bpexportglobals: TRUE; bpexportvariables: FALSE; bpforceGC: FALSE
##   bpfallback: TRUE
##   bplogdir: NA
##   bpresultdir: NA
##   cluster type: FORK

Arguments are described on the corresponding help page, e.g., ?MulticoreParam..

register()ing BiocParallelParam instances

The list of registered BiocParallelParam instances represents the user’s preferences for different types of back-ends. Individual algorithms may specify a preferred back-end, and different back-ends maybe chosen when parallel evaluation is nested.

The registry behaves like a ‘stack’ in that the last entry registered is added to the top of the list and becomes the “next used” (i.e., the default).

registered invoked with no arguments lists all back-ends.

registered()

## $MulticoreParam
## class: MulticoreParam
##   bpisup: FALSE; bpnworkers: 2; bptasks: 0; bpjobname: BPJOB
##   bplog: FALSE; bpthreshold: INFO; bpstopOnError: TRUE
##   bpRNGseed: ; bptimeout: NA; bpprogressbar: FALSE
##   bpexportglobals: TRUE; bpexportvariables: FALSE; bpforceGC: FALSE
##   bpfallback: TRUE
##   bplogdir: NA
##   bpresultdir: NA
##   cluster type: FORK
## 
## $SnowParam
## class: SnowParam
##   bpisup: FALSE; bpnworkers: 2; bptasks: 0; bpjobname: BPJOB
##   bplog: FALSE; bpthreshold: INFO; bpstopOnError: TRUE
##   bpRNGseed: ; bptimeout: NA; bpprogressbar: FALSE
##   bpexportglobals: TRUE; bpexportvariables: TRUE; bpforceGC: FALSE
##   bpfallback: TRUE
##   bplogdir: NA
##   bpresultdir: NA
##   cluster type: SOCK
## 
## $SerialParam
## class: SerialParam
##   bpisup: FALSE; bpnworkers: 1; bptasks: 0; bpjobname: BPJOB
##   bplog: FALSE; bpthreshold: INFO; bpstopOnError: TRUE
##   bpRNGseed: ; bptimeout: NA; bpprogressbar: FALSE
##   bpexportglobals: FALSE; bpexportvariables: FALSE; bpforceGC: FALSE
##   bpfallback: FALSE
##   bplogdir: NA
##   bpresultdir: NA

bpparam returns the default from the top of the list.

bpparam()

## class: MulticoreParam
##   bpisup: FALSE; bpnworkers: 2; bptasks: 0; bpjobname: BPJOB
##   bplog: FALSE; bpthreshold: INFO; bpstopOnError: TRUE
##   bpRNGseed: ; bptimeout: NA; bpprogressbar: FALSE
##   bpexportglobals: TRUE; bpexportvariables: FALSE; bpforceGC: FALSE
##   bpfallback: TRUE
##   bplogdir: NA
##   bpresultdir: NA
##   cluster type: FORK

Add a specialized instance with register. When default is TRUE, the new instance becomes the default.

default <- registered()
register(BatchtoolsParam(workers = 10), default = TRUE)

BatchtoolsParam has been moved to the top of the list and is now the default.

names(registered())

## [1] "BatchtoolsParam" "MulticoreParam"  "SnowParam"       "SerialParam"

bpparam()

## class: BatchtoolsParam
##   bpisup: FALSE; bpnworkers: 10; bptasks: 0; bpjobname: BPJOB
##   bplog: FALSE; bpthreshold: INFO; bpstopOnError: TRUE
##   bpRNGseed: NA; bptimeout: NA; bpprogressbar: FALSE
##   bpexportglobals: TRUE; bpexportvariables: TRUE; bpforceGC: FALSE
##   bpfallback: TRUE
##   bplogdir: NA
##   bpresultdir: NA
##   cluster type: multicore
##   template: NA
##   registryargs:
##     file.dir: /tmp/Rtmp3GgYoJ/Rbuild17e679dc6531/BiocParallel/vignettes/file1b7c1191490b
##     work.dir: getwd()
##     packages: character(0)
##     namespaces: character(0)
##     source: character(0)
##     load: character(0)
##     make.default: FALSE
##   saveregistry: FALSE
##   resources:

Restore the original registry

for (param in rev(default))
    register(param)

Parallel looping, vectorized and aggregate operations

These are used in common functions, implemented as much as possible for all back-ends. The functions (see the help pages, e.g., ?bplapply for a full definition) include

bplapply(X, FUN, ...):

Apply in parallel a function FUN to each element of X. bplapply invokes FUN length(X) times, each time with a single element of X.

bpmapply(FUN, ...):

Apply in parallel a function to the first, second, etc., elements of each argument in ….

bpiterate(ITER, FUN, ...):

Apply in parallel a function to the output of function ITER. Data chunks are returned by ITER and distributed to parallel workers along with FUN. Intended for iteration though an undefined number of data chunks (i.e., records in a file).

bpvec(X, FUN, ...):

Apply in parallel a function FUN to subsets of X.bpvec invokes function as many times as there are cores or cluster nodes, with receiving a subset (typically more than 1 element, in contrast to bplapply) of X.

bpaggregate(x, data, FUN, ...):

Use the formula in X to aggregate data using FUN.

Parallel evaluation environment

These functions query and control the state of the parallel evaluation environment.

bpisup(x): Query a BiocParallelParam back-end X for its status.

bpworkers; bpnworkers: Query a BiocParallelParam back-end for the number of workers available for parallel evaluation.

bptasks: Divides a job (e.g., single call to *lapply function) into tasks. Applicable to MulticoreParam only;DoparParam and BatchtoolsParam have their own approach to dividing a job among workers.

bpstart(x): Start a parallel back end specified by BiocParallelParam x,, if possible.

bpstop(x): Stop a parallel back end specified by BiocParallelParam x.

Error handling and logging

Logging and advanced error recovery is available in BiocParallel 1.1.25 and later. For a more details see the vignette titled “Error Handling and Logging”:

browseVignettes("BiocParallel")

Locks and counters

Inter-process (i.e., single machine) locks and counters are supported using ipclock(), ipcyield(), and friends. Use these to synchronize computation, e.g., allowing only a single process to write to a file at a time.

Forked processes with MulticoreParam

This example counts overlaps between BAM files and a defined set of ranges. First create a GRanges with regions of interest (in practice this could be large).

library(GenomicAlignments) ## for GenomicRanges and readGAlignments()
gr <- GRanges("chr14", IRanges((1000:3999)*5000, width=1000))

A ScanBamParam defines regions to extract from the files.

param <- ScanBamParam(which=range(gr))

FUN counts overlaps between the ranges in ‘gr’ and the files.

FUN <- function(fl, param) {
    gal <- readGAlignments(fl, param = param)
    sum(countOverlaps(gr, gal))
}

All parameters necessary for running a job in a multi-core environment are specified in the MulticoreParam instance.

MulticoreParam()

## class: MulticoreParam
##   bpisup: FALSE; bpnworkers: 2; bptasks: 0; bpjobname: BPJOB
##   bplog: FALSE; bpthreshold: INFO; bpstopOnError: TRUE
##   bpRNGseed: ; bptimeout: NA; bpprogressbar: FALSE
##   bpexportglobals: TRUE; bpexportvariables: FALSE; bpforceGC: FALSE
##   bpfallback: TRUE
##   bplogdir: NA
##   bpresultdir: NA
##   cluster type: FORK

The BiocParallel functions, such as bplapply, use information in the MulticoreParam to set up the appropriate back-end and pass relevant arguments to low-level functions.

> bplapply(fls[1:3], FUN, BPPARAM = MulticoreParam(), param = param)
$ERR127306
[1] 1185

$ERR127307
[1] 1123

$ERR127308
[1] 1241

Shared memory environments eliminate the need to pass large data between workers or load common packages. Note that in this code the GRanges data was not passed to all workers in bplapply and FUN did not need to load [GenomicAlignmentsfor access to the readGAlign ments function.

Problems with forked processes occur when code implementating functionality used by the workers is not written in anticipation of use by forked processes. One example is the database connection underlying Bioconductor’s org.* packages. This pseudo-code

library(org.Hs.eg.db)
FUN <- function(x, ...) {
...
mapIds(org.Hs.eg.db, ...)
...
}
bplapply(X, FUN, ..., BPPARAM = MulticoreParam())

is likely to fail, because library(org.Hs.eg.db) opens a database connection that is accessed by multiple processes. A solution is to ensure that the database is opened independently in each process

FUN <- function(x, ...) {
library(org.Hs.eg.db)
...
mapIds(org.Hs.eg.db, ...)
...
}
bplapply(X, FUN, ..., BPPARAM = MulticoreParam())

Clusters of independent processes with SnowParam

Both Windows and non-Windows machines can use the cluster approach to spawn processes. BiocParallel back-end choices for clusters on a single machine are SnowParam for configuring a Snow cluster or the DoparParam for use with the foreach package.

To re-run the counting example, FUN needs to modified such that ‘gr’ is passed as a formal argument and required libraries are loaded on each worker. (In general, this is not necessary for functions defined in a package name space, see Section 6.)

FUN <- function(fl, param, gr) {
    suppressPackageStartupMessages({
        library(GenomicAlignments)
    })
    gal <- readGAlignments(fl, param = param)
    sum(countOverlaps(gr, gal))
}

Define a 2-worker SOCK Snow cluster.

snow <- SnowParam(workers = 2, type = "SOCK")

A call to bplapply with the SnowParam creates the cluster and distributes the work.

bplapply(fls[1:3], FUN, BPPARAM = snow, param = param, gr = gr)

## $ERR127306
## [1] 1185
## 
## $ERR127307
## [1] 1123
## 
## $ERR127308
## [1] 1241

The FUN written for the cluster adds some overhead due to the passing of the GRanges and the loading of GenomicAlignments on each worker. This approach, however, has the advantage that it works on most platforms and does not require a coding change when switching between windows and non-windows machines.

If several bplapply() statements are likely to require the same resource, it often makes sense to create a cluster once using bpstart(). The workers are re-used by each call to bplapply(), so they do not have to re-load packages, etc.

register(SnowParam()) # default evaluation
bpstart() # start the cluster
...
bplapply(X, FUN1, ...)
...
bplapply(X, FUN2, ...) # re-use workers
...
bpstop()

Ad hoc Sockets

On Linux and Mac OS X, a socket cluster is created across machines by supplying machine names as the`workers``argument to a BiocParallelParam instance instead of a number. Each name represents an R process; repeat names indicate multiple workers on the same machine.

Create a with SnowParam 2 cpus from ‘rhino01’ and 1 from ‘rhino02’.

hosts <- c("rhino01", "rhino01", "rhino02")
param <- SnowParam(workers = hosts, type = "SOCK")

Execute FUN 4 times across the workers.

> FUN <- function(i) system("hostname", intern=TRUE)
> bplapply(1:4, FUN, BPPARAM = param)
[[1]]
[1] "rhino01"

[[2]]
[1] "rhino01"

[[3]]
[1] "rhino02"

[[4]]
[1] "rhino01"

When creating a cluster across Windows machines must be IP addresses (e.g., "140.107.218.57") instead of machine names.

### MPI

An MPI cluster across machines is created with mpirun or mpiexec from the command line or a script. A list of machine names provided as the -hostfile argument defines the mpi universe.

The hostfile requests 2 processors on 3 different machines.

rhino01 slots=2
rhino02 slots=2
rhino03 slots=2

From the command line, start a single interactive process on the current machine.

mpiexec --np 1 --hostfile hostfile R --vanilla

Load BiocParallel and create an MPI Snow cluster. The number workers of in should match the number of slots requested in the hostfile. Using a smaller number of workers uses a subset of the slots.

> library(BiocParallel)
> param <- SnowParam(workers = 6, type = "MPI")

Execute FUN 6 times across the workers.

> FUN <- function(i) system("hostname", intern=TRUE)
> bplapply(1:6, FUN, BPPARAM = param)
bplapply(1:6, FUN, BPPARAM = param)
[[1]]
[1] "rhino01"

[[2]]
[1] "rhino02"

[[3]]
[1] "rhino02"

[[4]]
[1] "rhino03"

[[5]]
[1] "rhino03"

[[6]]
[1] "rhino01"

Batch jobs can be launched with mpiexec and R CMD BATCH. Code to be executed is in ‘Rcode.R’.

mpiexec --hostfile hostfile R CMD BATCH Rcode.R

Cluster-centric

The idea is to use cluster management software to allocate resources, and then arrange for an script to be evaluated in the context of allocated resources. NOTE: Depending on your cluster configuration it may be necessary to add a line to the template file instructing workers to use the version of R on the master / head node. Otherwise the default R on the worker nodes will be used.

For SLURM, we might request space for 4 tasks (with salloc or sbatch), arrange to start the MPI environment (with orterun) and on a single node in that universe run an script BiocParallel-MPI.R. The command is

$ salloc -N 4 orterun -n 1 R -f BiocParallel-MPI.R

The R script might do the following, using MPI for parallel evaluation. Start by loading necessary packages and defining FUN work to be done

library(BiocParallel)
library(Rmpi)
FUN <- function(i) system("hostname", intern=TRUE)

Create a SnowParam instance with the number of nodes equal to the size of the MPI universe minus 1 (let one node dispatch jobs to workers), and register this instance as the default

param <- SnowParam(mpi.universe.size() - 1, "MPI")
register(param)

Evaluate the work in parallel, process the results, clean up, and quit

xx <- bplapply(1:100, FUN)
table(unlist(xx))
mpi.quit()

The entire session is as follows:

$ salloc -N 4 orterun -n 1 R --vanilla -f BiocParallel-MPI.R
salloc: Job is in held state, pending scheduler release
salloc: Pending job allocation 6762292
salloc: job 6762292 queued and waiting for resources
salloc: job 6762292 has been allocated resources
salloc: Granted job allocation 6762292
## ...
> FUN <- function(i) system("hostname", intern=TRUE)
>
> library(BiocParallel)
> library(Rmpi)
> param <- SnowParam(mpi.universe.size() - 1, "MPI")
> register(param)
> xx <- bplapply(1:100, FUN)
> table(unlist(xx))
gizmof13 gizmof71 gizmof86 gizmof88
25 25 25 25
>
> mpi.quit()
salloc: Relinquishing job allocation 6762292
salloc: Job allocation 6762292 has been revoked.

One advantage of this approach is that the responsibility for managing the cluster lies firmly with the cluster management software – if one wants more nodes, or needs special resources, then adjust parameters to salloc (or sbatch).

Notice that workers are spawned within the bplapply function; it might often make sense to more explicitly manage workers with bpstart and bpstop, e.g.,

param <- bpstart(SnowParam(mpi.universe.size() - 1, "MPI"))
register(param)
xx <- bplapply(1:100, FUN)
bpstop(param)
mpi.quit()

R-centric

A more R-centric approach might start an R script on the head node, and use batchtools to submit jobs from within R the session. One way of doing this is to create a file containing a template for the job submission step, e.g., for SLURM; a starting point might be found at

tmpl <- system.file(package="batchtools", "templates", "slurm-simple.tmpl")
noquote(readLines(tmpl))

##  [1] #!/bin/bash                                                                                                 
##  [2]                                                                                                             
##  [3] ## Job Resource Interface Definition                                                                        
##  [4] ##                                                                                                          
##  [5] ## ntasks [integer(1)]:       Number of required tasks,                                                     
##  [6] ##                            Set larger than 1 if you want to further parallelize                          
##  [7] ##                            with MPI within your job.                                                     
##  [8] ## ncpus [integer(1)]:        Number of required cpus per task,                                             
##  [9] ##                            Set larger than 1 if you want to further parallelize                          
## [10] ##                            with multicore/parallel within each task.                                     
## [11] ## walltime [integer(1)]:     Walltime for this job, in seconds.                                            
## [12] ##                            Must be at least 60 seconds for Slurm to work properly.                       
## [13] ## memory   [integer(1)]:     Memory in megabytes for each cpu.                                             
## [14] ##                            Must be at least 100 (when I tried lower values my                            
## [15] ##                            jobs did not start at all).                                                   
## [16] ##                                                                                                          
## [17] ## Default resources can be set in your .batchtools.conf.R by defining the variable                         
## [18] ## 'default.resources' as a named list.                                                                     
## [19]                                                                                                             
## [20] <%                                                                                                          
## [21] # relative paths are not handled well by Slurm                                                              
## [22] log.file = fs::path_expand(log.file)                                                                        
## [23] -%>                                                                                                         
## [24]                                                                                                             
## [25]                                                                                                             
## [26] #SBATCH --job-name=<%= job.name %>                                                                          
## [27] #SBATCH --output=<%= log.file %>                                                                            
## [28] #SBATCH --error=<%= log.file %>                                                                             
## [29] #SBATCH --time=<%= ceiling(resources$walltime / 60) %>                                                      
## [30] #SBATCH --ntasks=1                                                                                          
## [31] #SBATCH --cpus-per-task=<%= resources$ncpus %>                                                              
## [32] #SBATCH --mem-per-cpu=<%= resources$memory %>                                                               
## [33] <%= if (!is.null(resources$partition)) sprintf(paste0("#SBATCH --partition='", resources$partition, "'")) %>
## [34] <%= if (array.jobs) sprintf("#SBATCH --array=1-%i", nrow(jobs)) else "" %>                                  
## [35]                                                                                                             
## [36] ## Initialize work environment like                                                                         
## [37] ## source /etc/profile                                                                                      
## [38] ## module add ...                                                                                           
## [39]                                                                                                             
## [40] ## Export value of DEBUGME environemnt var to slave                                                         
## [41] export DEBUGME=<%= Sys.getenv("DEBUGME") %>                                                                 
## [42]                                                                                                             
## [43] <%= sprintf("export OMP_NUM_THREADS=%i", resources$omp.threads) -%>                                         
## [44] <%= sprintf("export OPENBLAS_NUM_THREADS=%i", resources$blas.threads) -%>                                   
## [45] <%= sprintf("export MKL_NUM_THREADS=%i", resources$blas.threads) -%>                                        
## [46]                                                                                                             
## [47] ## Run R:                                                                                                   
## [48] ## we merge R output with stdout from SLURM, which gets then logged via --output option                     
## [49] Rscript -e 'batchtools::doJobCollection("<%= uri %>")'

The R script, run interactively or from the command line, might then look like

## define work to be done
FUN <- function(i) system("hostname", intern=TRUE)
library(BiocParallel)

## register SLURM cluster instructions from the template file
param <- BatchtoolsParam(workers=5, cluster="slurm", template=tmpl)
register(param)

## do work
xx <- bplapply(1:100, FUN)
table(unlist(xx))

The code runs on the head node until bplapply , where the script interacts with the SLURM scheduler to request a SLURM allocation, run jobs, and retrieve results. The argument 4 to BatchtoolsParam specifies the number of workers to request from the scheduler; bplapply divides the 100 jobs among the 4 workers. If BatchtoolsParam had been created without specifying any workers, then 100 jobs implied by the argument to bplapply would be associated with 100 tasks submitted to the scheduler.

Because cluster tasks are running in independent R instances, and often on physically separate machines, a convenient ‘best practice’ is to write FUN in a ‘functional programming’ manner, such that all data required for the function is passed in as arguments or (for large data) loaded implicitly or explicitly (e.g., via an R library) from disk.