Socket clusters

Another important component of slurmR is makeSlurmCluster function. This allow users creating multi-node PSOCKCluster class objects. The implementation of this function, wrapper of parallel::makePSOCKcluster, is very simple:

1. It submits a job to Slurm requesting the desired number of tasks. Each task will then return information regarding the node at which it is operating.

2. Once Slurm allocates the resources, the master R session (from which the job was submitted) will read in the node names returned by each task.

3. With the full list of nodenames in usage, makeSlurmCluster will pass the list of names to parallel::makePSOCKcluster, which ultimately creates the cluster class object.

After creating the cluster object, the workflow is exactly the same as with the parallel package. Here is an example from the makeSlurmCluster manual

# Creating a cluster with 100 workers/offpring/child R sessions
cl <- makeSlurmCluster(100)

# Computing the mean of a 100 random uniforms within each worker
# for this we can use any of the function available in the parallel package.
ans <- parSapply(1:200, function(x) mean(runif(100)))

# We simply call stopCluster as we would do with any other cluster
# object
stopCluster(ans)

Job arrays using the *apply family

Whenever Slurm_lapply, Slurm_sapply, or Slurm_Map are called, a lot of things happen under the hood. What the user does not see is the way in which slurmR sets us a job and submits it to the queue.

Just like rslurm, slurmR has two levels of job distribution: first, Slurm Jobs, and second, within each job via parallel::mclapply and parallel::mcMap (task forking). In general, the function Slurm_* is implemented as follows:

1. List whatever R packages are loaded, including the path to the R package.

2. List all the objects passed via ellipsis (...), and, together with X and FUN or f, save them at [tmp_path]/[job_name]/ as [object-name].rds.

3. Write out the corresponding R script and Slurm bash file, and save them as [tmp_path]/[job_name]/00-rscript.r, and [tmp_path]/[job_name]/01-bash.sh respectively.

4. If plan = "collect" (the default), the job will be submitted to the queue via sbatch(), and the function will wait until is flagged as completed by Slurm.

5. Once sbatch() is called, a Job Array will be submitted in which each R job will lunch up to mc.cores forked processes (second layer of palatalization)

Once it is done, the the results can be collected using Slurm_collect, which happens automatically if the user set plan = "collect".

The next section discusses some advantages of submitting jobs using socket clusters versus job arrays.

Sockets vs Arrays

While socket clusters, created via makePSOCKcluster or, in the case of slurmR, via makeSlurmCluster, may be more efficient in terms of data communication¹, using job arrays has some important benefits over socket cluster:

1. The number of workers can be much higher than clusters with the parallel package.² Users needing to work with hundreds or thousands of jobs/instances may need to use job arrays instead.

2. If part of the job fails due to a failure of one of the tasks in the array, the job can be easily resubmitted. The same is not necessarily true for socket clusters.

3. Job arrays can run independently from the main session that started the job. This means that, if for some reason the main session crashes or stops, the job arrays will continue working regardless, and what’s more, the results can be collected anyway.

Single node (machine), multi-core simulation

Using parallel::mclapply, we could just type

set.seed(12)
ans <- parallel::mclapply(rep(1e6, 100), simpi)
mean(unlist(ans))

Which estimates pi using a single node(computer).

Multi-job submission with job arrays

In the case of job arrays, we can use the Slurm_lapply function implemented in the package. Before submitting a job to the queue, we need to specify some options that are needed to create it:

• tmp_path: A path to a directory to which all computing nodes of the cluster have read+write access.

• job_name: The name of the job, passed to sbatch via the job-name flag. This will also be used as the name of the folder that is created within tmp_path.

Ultimately, all the objects saved by the job will be located in the path defined by tmp_path/job_name.

library(slurmR)

# Setting required parameters
opts_slurmR$set_tmp_path("/stagging/slurmr-jobs/")
opts_slurmR$set_job_name("simulating-pi")

Moreover, we can specify more options to be set as default options for all the jobs submitted for the current session. For example, we can set the default partition and account as follows:

# Optional parameters are set via set_opts
opts_slurmR$set_opts(partition="conti", account="lc_dvc")

A comprehensive list of options can be found here. To see what are all the current defaults, we can just print the opts_slurmR object:

opts_slurmR

## 
## Options for sbatch (Slurm workflow):
##   partition   : conti
##   account     : lc_dvc
##   job-name    : simulating-pi
## 
## Preamble:
##   n/a
## 
## Other options (R workflow):
##   tmp_path    : /stagging/slurmr-jobs/
##   cmd         : sbatch
##   verbose     : FALSE
##   debug       : FALSE
## 
## To get and set options for Slurm jobs creation use (see ?opts_slurmR):
## 
## debug_off : function ()  
## debug_on : function ()  
## get_cmd : function ()  
## get_debug : function ()  
## get_job_name : function (check = TRUE)  
## get_opts_job : function (...)  
## get_opts_r : function (...)  
## get_preamble : function ()  
## get_tmp_path : function ()  
## get_verbose : function ()  
## reset : function ()  
## set_job_name : function (name)  
## set_opts : function (...)  
## set_preamble : function (...)  
## set_tmp_path : function (path = Sys.getenv("SLURMR_TMP_PATH", getwd()))  
## verbose_off : function ()  
## verbose_on : function ()

Once we have specified all the needed options, we can do our Slurm_lapply call and submit the job to the queue as follows:

job <- Slurm_lapply(rep(1e6, 100), simpi, njobs=10, mc.cores=10, plan = "wait")

If plan = "wait", then Slurm_lapply will return once the job is done (or failed). To collect the results we can use the Slurm_collect function:

ans <- Slurm_collect(job)
mean(unlist(ans))

Alternatively, we could have collected the results on the fly by telling slurmR that the plan is to "collect" the results:

ans <- Slurm_lapply(rep(1e6, 100), simpi, njobs=10, mc.cores=10, plan = "collect")
mean(unlist(job))

This way Slurm_lapply will do the Slurm_collect call before returning.

Multi-node cluster object

Another way to do this is using parallel::parLapply with a multi-node socket cluster.³ To do this, we can use the makeSlurmCluster function and proceed as follows:

cl <- makeSlurmCluster(50)
ans <- parallel::parLapply(cl, rep(1e6, 100), simpi)
mean(unlist(ans))

Once we are done with the calculations, we can stop the cluster object by simply calling the stopCluster function:

stopCluster(cl)

And slurmR will kill the job (and thus, the socket connections) calling scancel.