How to implement parallel jags on Windows with foreach?

Question

I would like to run jags models in parallel on my windows computer with 4 cores, but have not been able to figure out why my model will not run. I have searched the web extensively including these posts:

http://andrewgelman.com/2011/07/23/parallel-jags-rngs/

http://users.soe.ucsc.edu/~draper/eBay-Google-2013-parallel-rjags-example.txt

When I run a simple example (see code below) with %do%, the model runs fine (serially of course). When I use %dopar%, I receive the error: Error in { : task 1 failed - "Symbol table is empty"

library(rjags)
library(coda)
library(foreach)              
library(doParallel)
library(random)     
load.module("lecuyer")

###  Data generation 
y <- rnorm(100)
n <- length(y)
win.data <- list(y=y, n=n)

# Define model
sink("model.txt")
cat("
    model {
    # Priors
    mu ~ dnorm(0, 0.001)
    tau <- 1 / (sigma * sigma)
    sigma ~ dunif(0, 10)
    # Likelihood
    for (i in 1:n) {
      y[i] ~ dnorm(mu, tau)
    }
}
",fill=TRUE)
sink()

inits <- function(){ list(mu=rnorm(1), sigma=runif(1, 0, 10),
                 .RNG.name = "lecuyer::RngStream", 
                 .RNG.seed = as.numeric(randomNumbers( n = 1, min = 1, max = 1e+06, col = 1 )) ) }
params <- c('mu','sigma')  

cl <- makePSOCKcluster(3)              
clusterSetRNGStream(cl)
registerDoParallel(cl)      
model.wd <- paste(getwd(), '/model.txt', sep='')     # I wondered if the cores were having trouble finding the model.         

m <- foreach(i=1:3, .packages=c('rjags','random','coda'), .multicombine=TRUE) %dopar% {  
                load.module( "lecuyer" )  
                model.jags <- jags.model(model.wd, win.data, inits=inits, n.chains=1, n.adapt=1000, quiet=TRUE)
                result <- coda.samples(model.jags, params, 1000, thin=5)
                return(result)
              }            
stopCluster(cl)
# Error in { : task 1 failed - "Symbol table is empty


sessionInfo()
# R version 3.0.1 (2013-05-16)
# Platform: x86_64-w64-mingw32/x64 (64-bit)
# 
# locale:
#   [1] LC_COLLATE=English_Canada.1252  LC_CTYPE=English_Canada.1252    LC_MONETARY=English_Canada.1252
# [4] LC_NUMERIC=C                    LC_TIME=English_Canada.1252    
# 
# attached base packages:
#   [1] parallel  stats     graphics  grDevices utils     datasets  methods   base     
# 
# other attached packages:
#   [1] random_0.2.1     doParallel_1.0.3 iterators_1.0.6  foreach_1.4.1    rjags_3-10       coda_0.16-1     
# [7] lattice_0.20-21 
# 
# loaded via a namespace (and not attached):
#   [1] codetools_0.2-8 compiler_3.0.1  grid_3.0.1      tools_3.0.1  

---

More Details:

The problem occurs on a Windows 7 computer with NO admin privaleges, but not on a computer WITH admin privaleges. The problem occurs with Rgui and Rterm and with the new rjags packaged 3-11. The error message occurs within the function jags.model

The problem appears to stem from a mismatch in writing and reading files to a temporary directory. When I start R, it automatically creates a temporary folder. When I close R, this folder is automatically deleted, unless it contains files.

For example, when I start R it creates this folder: C:\Users\jesse whittington\AppData\Local\Temp\RtmpoBe1gw.

When I run a rjags model with

m <- jags.model(file='model.txt', data=win.data, inits=inits, n.chains=3, n.adapt=1000, quiet=FALSE)

No files are written to this temporary directory.

When I run 3 chains serially with foreach and %do%, 3 temporary files are written to this folder. These files are 1 kb in size and when I open with a text editor they appear blank.

wd <- getwd()                   
cl <- makePSOCKcluster(3, outfile=paste(wd,'/Out_messages.txt', sep=''))   # 3 chains           
clusterSetRNGStream(cl)
registerDoParallel(cl)   
m <- foreach(i=1:3, .packages=c('rjags','random','coda'), .multicombine=TRUE) %do% {  
                load.module( "lecuyer" ) 
                result <- jags.model(file='model.txt', data=win.data, inits=inits, n.chains=1, n.adapt=1000, quiet=FALSE)
                return(result)
              }  
stopCluster(cl) 

When I run 3 chains in parallel with foreach and %dopar%, 3 temporary files are written to the folder ..Temp\RtmpoBe1gw. The error messages in the outfile suggest that the function is looking for DIFFERENT files in DIFFERENT temporary directories. When, I include a line to create a tempfile directory and name, I see that 3 new temporary folders are created (they are later deleted with stopCluster). jags.model looks in these 3 folders for the temporary files and fails because there is nothing in them. Thus, I suspect tempfiles are written to one temporary directory (associated with the parent R session) and then fails when trying to open different tmpfiles in the 3 temporary directories created within foreach.

wd <- getwd()                   
cl <- makePSOCKcluster(3, outfile=paste(wd,'/Out_messages.txt', sep=''))   # 3 chains           
clusterSetRNGStream(cl)
registerDoParallel(cl)   
m <- foreach(i=1:3, .packages=c('rjags','random','coda'), .multicombine=TRUE) %dopar% {  
                load.module( "lecuyer" ) 
        tmp <- tempfile()
                print(tmp)
                result <- jags.model(file='model.txt', data=win.data, inits=inits, n.chains=1, n.adapt=1000, quiet=FALSE)
                return(result)
              }  
stopCluster(cl) 

From Out_messages.txt

starting worker pid=4396 on localhost:11109 at 08:34:06.430
starting worker pid=6548 on localhost:11109 at 08:34:06.879
starting worker pid=6212 on localhost:11109 at 08:34:07.418
Loading required package: coda
Loading required package: lattice
Loading required package: coda
Loading required package: lattice
Loading required package: coda
Loading required package: lattice
Linked to JAGS 3.3.0
Loaded modules: basemod,bugs
Linked to JAGS 3.3.0
Loaded modules: basemod,bugs
Linked to JAGS 3.3.0
Loaded modules: basemod,bugs
module lecuyer loaded
module lecuyer loaded
module lecuyer loaded
[1] "C:\\Users\\JESSEW~1\\AppData\\Local\\Temp\\RtmpQbPAVC\\file112c8077a0"  # Note this is from: tmp <- tempfile()
[1] "C:\\Users\\JESSEW~1\\AppData\\Local\\Temp\\RtmpMPMpcY\\file199489564c6"
[1] "C:\\Users\\JESSEW~1\\AppData\\Local\\Temp\\Rtmpk9vMR5\\file18445f6b2fd4"
Compiling model graph
Compiling model graph
Compiling model graph

Warning messages:
1: In jags.model(file = "model.txt", data = win.data, inits = inits,  :
  Unused variable "y" in data
2: In jags.model(file = "model.txt", data = win.data, inits = inits,  :
  Unused variable "n" in data
3: In jags.model(file = "model.txt", data = win.data, inits = inits,  :
  Failed to open file C:\Users\JESSEW~1\AppData\Local\Temp\RtmpQbPAVC\file112c394b4eef
Nothing to compile

4: In jags.model(file = "model.txt", data = win.data, inits = inits,  :
  Unused initial value for "mu" in chain 1
5: In jags.model(file = "model.txt", data = win.data, inits = inits,  :
  Unused initial value for "sigma" in chain 1
6: In jags.model(file = "model.txt", data = win.data, inits = inits,  :
  Can't initialize. No nodes in graph (Have you compiled the model?)

The folder RtmpQbPAVC is created but the file file112c394b4eef does not exist.

Answer 1

I have identified the source of the problem. I can write and read files to and from a temporary directory when using R normally. When in parallel, I can write files to the temporary directories, but I do NOT have permission to read files.

The problem occurs both writing and reading text files (using writeLines and readLines) and csv files.

I have since found that if I receive this message: "Error in { : task 1 failed - cannot open the connection", I can rectify the problem by deleting all temporary files in TEMP. For some locked files, I have to shut down and restart the computer before I am able to delete the necessary files. Even so, within the same R session I might receive the error message and then be able to successful run the program on my next try. The problem likely stems from our government anti-virus software and/or the structure of our remote network access.

Here is an example that writes and reads text files for simplicity.

library(foreach)              
library(doParallel)
wd <- getwd()
data <- data.frame(x=1:10, y=1:10)

This works fine.

modfile <- tempfile()
print(modfile)
# "C:\\Users\\JESSEW~1\\AppData\\Local\\Temp\\RtmpsvYfFk\\filef38a272022"
write.csv(data, modfile, row.names=F)
m <- read.csv(modfile) 

This does not work

cl <- makePSOCKcluster(3, outfile=paste(wd,'/Out_messages.txt', sep=''))   # 3 chains           
clusterSetRNGStream(cl)
registerDoParallel(cl)   
m <- foreach(i=1:3) %dopar% {  
  modfile <- tempfile()
  write.csv(data, modfile, row.names=F)
  x <- read.csv(modfile)
  return(x)
}  
# Error in { : task 1 failed - "cannot open the connection"
stopCluster(cl) 

Here is the output from Out_message.txt. Note the "Permission Denied" on the far right.

starting worker pid=6852 on localhost:11611 at 22:09:19.488
starting worker pid=6984 on localhost:11611 at 22:09:19.926
starting worker pid=3384 on localhost:11611 at 22:09:20.441
Warning message:
  Warning message:
  In file(con, "r") :
  cannot open file 'C:\Users\JESSEW~1\AppData\Local\Temp\Rtmp6dEZLP\file1ac44a506032': Permission denied
In file(con, "r") :
  cannot open file 'C:\Users\JESSEW~1\AppData\Local\Temp\RtmpuydRvR\file1b48185f2a2d': Permission denied
Warning message:
  In file(con, "r") :
  cannot open file 'C:\Users\JESSEW~1\AppData\Local\Temp\RtmpAbOIng\filed382ef37d51': Permission denied

Answer 2

Since the errors are caused by writing and reading the model file, I suggest that you bypass that issue by using the "textConnection" function. This can be used to create a file-like object without creating an actual file, thus avoiding the need for temporary files. I modified your example to demonstrate this:

library(rjags)
library(doParallel)
library(random)
load.module("lecuyer")
y <- rnorm(100)
n <- length(y)
win.data <- list(y=y, n=n)
model <- "
  model {
    # Priors
    mu ~ dnorm(0, 0.001)
    tau <- 1 / (sigma * sigma)
    sigma ~ dunif(0, 10)
    # Likelihood
    for (i in 1:n) {
      y[i] ~ dnorm(mu, tau)
    }
  }"
inits <- function() {
  list(mu=rnorm(1), sigma=runif(1, 0, 10),
       .RNG.name="lecuyer::RngStream",
       .RNG.seed=as.numeric(randomNumbers(n=1, min=1, max=1e+06, col=1)))
}
params <- c('mu', 'sigma')
cl <- makePSOCKcluster(3)
clusterSetRNGStream(cl)
registerDoParallel(cl)

m <- foreach(i=1:3, .packages=c('rjags', 'random'),
             .combine='c', .final=mcmc.list) %dopar% {
  load.module( "lecuyer" )
  model.jags <- jags.model(textConnection(model), win.data, inits=inits,
                           n.chains=1, n.adapt=1000, quiet=TRUE)
  coda.samples(model.jags, params, 1000, thin=5)
}

I also changed the result handling so that the value returned by the foreach loop is an "mcmc.list" object, which is what the "coda.samples" function returns.

Answer 3

Steve brought this to my attention, but your second example shows that it is not a problem with rjags. I am unable to reproduce the bug in either example using the same setup (Windows 7, R 3.0.1, JAGS 3.0.3, ordinary user without admin access).