Interactive distributed computing in R

Use case: interactive cluster computing from RStudio.

For batch jobs, Starcluster is likely a better choice. Brian Holt (Unlicensed) wrote up a document on using R in starcluster called /wiki/spaces/IT/pages/7867417. Other documents give an overview of Distributed Computation Strategy and how to run Distributed Compute Jobs.

The BioConductor group has put together a Cloud Formation stack for doing interactive parallel computing in R. Follow those instructions, selecting the number of workers and size of the EC2 instances. Once the stack comes up, which took about 10 minutes for me, you log into RStudio on the head node. You'll start R processes on the worker nodes and send commands to the workers.

Note that the parallel package is perfectly happy starting up several copies of R on a single machine, which can be helpful for testing.

Starting a cluster

The IP addresses of the workers (and the head node) get stored on the head node in a file. We'll read that file and create an R process for each core on each worker.

# grab host IPs
lines <- readLines("/usr/local/Rmpi/hostfile.plain")

# we'll want to start a worker for each core on each
# machine in the cluster
hosts <- do.call(c, lapply(strsplit(lines, " "), function(host) { rep(host[1], as.integer(host[2])) }))

library(parallel)
help(package=parallel)

cl <- makePSOCKcluster(hosts)

Simple tests

Try a few simple tests to make sure we're able to evaluate code on the workers and that it buys us some speed.

# try something simple
ans <- unlist(clusterEvalQ(cl, { mean(rnorm(1000)) }), use.names=F)

# test a time-consuming job
system.time(ans <- clusterEvalQ(cl, { sapply(1:1000, function(i) {mean(rnorm(10000))}) }))

# do the same thing locally
system.time(ans2 <- sapply(1:(1000*length(hosts)), function(i) {mean(rnorm(10000))}))

Head node vs. workers

Be aware of when you're running commands on the head node and when commands are running on the workers. Many commands will be better off running on the head node. When it's time to do something in parallel, you'll need to ship data objects to the workers, which is done with clusterExport, something like the following pattern:

myBigData <- computeBigDataMatrix(fizz, buzz)
moreData  <- constructPhatDataFrame(x, y, z)

clusterExport(cl, c('myBigData', 'moreData'))

results <- clusterEvalQ(cl, { for (bootstrap_runs in 1:10) { computeOn(myBigData, moreData) } })

Loading packages

The BioC virtual machine image comes with tons of good stuff already installed, but inevitably, you'll need to install something else.

It might be necessary to modify the library path. If you try to install packages on the workers and get an error to the effect that the workers "cannot install packages", you need to do this.

# set lib path to install packages

clusterEvalQ(cl, { .libPaths( c('/home/ubuntu/R/library', .libPaths()) ) })
clusterEvalQ(cl, {
    install.packages("someUsefulPackage")
    require(someUsefulPackage)
})

Loading sage package

clusterEvalQ(cl, {
    options(repos=structure(c(CRAN="http://cran.fhcrc.org/")))
    source('http://depot.sagebase.org/CRAN.R')

    pkgInstall("synapseClient")
    pkgInstall("predictiveModeling")
    
    library(synapseClient)
    library(predictiveModeling)
})

Loading synapse entities

Logging in.

clusterEvalQ(cl, { synapseLogin('joe.user@mydomain.com','secret') })

Asking many worker nodes to request Synapse entities at once is a fun and easy way to mount a distributed denial of service attack on the repository service. The service deals with this by timing out requests, which means some workers will succeed, while others will fail. A couple of tricks will help smooth over these problems.

check if our target data already exists. That way, we can re-try in the event of partial failure without re-doing work and unnecessarily thrashing Synapse.
throw in a few random seconds of rest for our workers. This spreads out the load on Synapse.

clusterEvalQ(cl, {
    if (!exists('expr')) {
        Sys.sleep(runif(1,0,5))
        expr_entity <- loadEntity('syn269056')
        expr <- expr_entity$objects$eSet_expr
    }
})

Accessing source code repos on worker nodes

Getting code onto the worker nodes can be done like so:

clusterEvalQ(cl, {
    system('svn export  --no-auth-cache --non-interactive --username joe.user --password supeRsecRet77 https://sagebionetworks.jira.com/svn/COMPBIO/trunk/users/juser/fantasticAnalysis.R')
})

<<github example>>

Return values

Return values from distributed computations have to come across a socket connection, so be careful what you return. Status values such as dim(result) can confirm that a computation succeeded and are often better than returning a whole result

clusterEvalQ(cl, {
    result <- produceGiantResultMatrix(foo, bar, bat)
    dim(result)
})

Also, consider putting intermediate values in synapse, which might serve as a means of checkpointing lengthy computations.

<<synapse example>>

Stopping a cluster

stopCluster(cl)

Don't forget to delete the stack in the AWS administration console to avoid continuing charges.

To do

These items will be added to this document, if and when I figure them out.

Spot instances? Is this worthwhile for interactive use?
Create our own Cloud Formation template
Run a user-specified script on start-up