fwrite hangs if called from within mclapply

[rdiaz02]

When calling fwrite from inside mclapply the execution hangs.

Here are two reproducible examples. The first uses a unique temporary file name via tempfile; the second doesn't.

library(data.table)
library(parallel)

mm <- matrix(1:10, ncol = 2)

f1 <- function(data) {
    tmpn <- tempfile()
    cat(paste("\n tmpn is ", tmpn, "\n"))
    write(t(data), file = tmpn)
}

f2 <- function(data) {
    tmpn <- tempfile()
    cat(paste("\n tmpn is ", tmpn, "\n"))
    data.table::fwrite(data.table(data), file = tmpn, verbose = TRUE)
}

## OK
f1(mm)
## OK
f2(mm)

## OK
mclapply(1:3, function(x) f1(mm), mc.silent = FALSE)
## OK; of course, just using lapply
mclapply(1:3, function(x) f2(mm), mc.silent = FALSE, mc.cores = 1)
## Hangs; does not even show attempting the third
mclapply(1:3, function(x) f2(mm), mc.silent = FALSE, mc.cores = 2)

Identical to the above, but without the tempfile (we create a filename using the index of mclapply), to make sure that is not the problem:

library(data.table)
library(parallel)

mm <- matrix(1:10, ncol = 2)

f1b <- function(index, data) {
    tmpn <- paste0("/tmp/afile", index)
    cat(paste("\n tmpn is ", tmpn, "\n"))
    write(t(data), file = tmpn)
}

f2b <- function(index, data) {
    tmpn <- paste0("/tmp/afile", index)
    cat(paste("\n tmpn is ", tmpn, "\n"))
    data.table::fwrite(data.table(data), file = tmpn, verbose = TRUE)
}

## OK
f1b(1, mm)
## OK
f2b(1, mm)

## OK
mclapply(1:3, function(x) f1b(index = x, data = mm), mc.silent = FALSE)
## OK; of course, just using lapply
mclapply(1:3, function(x) f2b(index = x, data = mm), mc.silent = FALSE, mc.cores = 1)
## Hangs and we do not even get to the third
mclapply(1:3, function(x) f2b(index = x, data = mm), mc.silent = FALSE, mc.cores = 2)

Expected behavior: run finishes without hanging, much like it does with write.

What does verbose in fwrite say? The runs show

tmpn is  /tmp/afile1 

 tmpn is  /tmp/afile2 
Maximum line length is 26 calculated in 0.000s
Writing column names ... done in 0.000s
Writing data rows in 1 batches of 40329 rows (each buffer size 1.000MB, turbo=1) ... Maximum line length is 26 calculated in 0.000s
Writing column names ... done in 0.000s
Writing data rows in 1 batches of 40329 rows (each buffer size 1.000MB, turbo=1) ... 

What can I see if I look at the stuff that is written?

If I look at afile1 and afile2, in the run that hangs, the only thing that is outputed is the header, not the contents of the file.

Output from sessionInfo()

> sessionInfo()
R version 3.3.0 Patched (2016-05-26 r70677)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Debian GNU/Linux stretch/sid

locale:
 [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
 [3] LC_TIME=en_GB.UTF-8        LC_COLLATE=en_US.UTF-8    
 [5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8   
 [7] LC_PAPER=en_GB.UTF-8       LC_NAME=C                 
 [9] LC_ADDRESS=C               LC_TELEPHONE=C            
[11] LC_MEASUREMENT=en_GB.UTF-8 LC_IDENTIFICATION=C       

attached base packages:
[1] parallel  stats     graphics  grDevices utils     datasets  methods  
[8] base     

other attached packages:
[1] data.table_1.9.7

loaded via a namespace (and not attached):
[1] compiler_3.3.0 tools_3.3.0   
> 

Is the cause of problem that the object we are writing is the same?
I would not expect that to be an issue, but that is not the problem, because this happens even if we write different R objects (here, each call modifies the object before writing it):

library(data.table)
library(parallel)

mm <- matrix(1:10, ncol = 2)

f1c <- function(index, data) {
    tmpn <- paste0("/tmp/afile", index)
    cat(paste("\n tmpn is ", tmpn, "\n"))
    data <- data[-index, ]
    write(t(data), file = tmpn)
}

f2c <- function(index, data) {
    tmpn <- paste0("/tmp/afile", index)
    cat(paste("\n tmpn is ", tmpn, "\n"))
    data <- data[-index, ]
    data.table::fwrite(data.table(data), file = tmpn, verbose = TRUE)
}

## OK
f1c(1, mm)
## OK
f2c(1, mm)

## OK
mclapply(1:3, function(x) f1c(index = x, data = mm), mc.silent = FALSE)
## OK; of course, just using lapply
mclapply(1:3, function(x) f2c(index = x, data = mm), mc.silent = FALSE, mc.cores = 1)
## Hangs and we do not even get to the third
mclapply(1:3, function(x) f2c(index = x, data = mm), mc.silent = FALSE, mc.cores = 2)

[arunsrinivasan]

Can't reproduce on OS X El Capitan, v1.9.7 latest.

[rdiaz02]

I just run it again (with data table downloaded right now), with same results, on two different Linux machines. I guess it might be an OS/compiler dependent issue?

[yitang]

@rdiaz02 I am having the same problem on my debian machine. apart from fwrite, i also noticed other functions in data.table that if used in mclapply, R hangs. The cpu usage goes down to and stay at 0.

[yitang]

I don't have the same problem on another old linux machine though. I guess it's probably caused by c libraries so I don't expect it will be resolved soon. I'd replace with fwrite with saveRDS at the moment.

[ChristK]

This sound like another case of implicit/explicit parallelisation conflict. Have you tried setthreads(1L) after loading data.table to force fwrite use only one thread?

[jangorecki]

@yitang Here is example of mixing parallel package and parallel functions in data.table, you should be able to measure performance and choose your working/preferred parallelism layer by setting two options.

[yitang]

@ChristK @jangorecki thank you very much. I didn't know setthreads function before. It would solve the problem. thanks.