[Bioc-devel] BiocParallel-devel error
Hi Valerie, Excellent. In addition to collecting log outputs, I have a few more suggestions that may be worth considering: - Collecting the results form parallel computing tasks directly in an R object is a great convenience, which I like a lot. However, in the context of slow computations there should be an option to redirect to files instead and then assemble things in R in a second step that the user can control. Perhaps this is possible already but it is not clear to me what the intended way is how to do this. - A much higher level of fault tolerance by adding options to restart failed jobs is another extremely important feature for parallel computations. This may only be possible if results are temporarily stored in files. For instance, if I farm out a computation to 10 compute nodes and one of them crashes, I want to be able to use the results form the 9 completed tasks but easily restart the computation assigned to the crashed node so that I get the final result quickly. BatchJobs provides most of these facilities. Making it easier and/or more obvious how to use these utilities from within BiocParallel may be all what is needed. Thomas
On Thu, Nov 20, 2014 at 04:43:54PM +0000, Valerie Obenchain wrote:
Interesting. I'm glad you figured that out. I'll add a line to the vignette that mentions this as a possible configuration issue. Using BatchJobs w/in BiocParallel should provide the same log files as using BatchJobs alone. I haven't tested it but my understanding was that the same options could be passed through '...' to name output files, directories etc. The other backends do not have logging ability. One of my current projects is to add logging to BiocParallel that works for all backends. I've been trying out the futile.logger package which looks promising for logging at both the system and user levels. We're also exploring more interactive debugging when an error is thrown. Valerie On 11/19/14 19:48, Thomas Girke wrote:
Hi Valerie, Michel and others, Finally, I freed up some time to revisit this problem. As it turns out, it is related to the use of a module system on our cluster. If I add in the template file for Torque (torque.tmpl) an explicit module load line for the specific R version, I am using on the master/head node, like this module load R/3.1.2-dev then everything runs as expected without errors in any of the more recent R release and development versions. Without this line the R version on the compute nodes will be the one used by default, which may result in an R version collision when submitting jobs from a different R version (e.g. R-dev). The reason that things worked in my specific case with BatchJobs but not with BiocParallel may simply be related to a less stringent enforcement of R version matches. Sorry that I didn't try this simple solution earlier. I guess what would help to isolate these kinds of problems in the future is a log file containing the STDOUTs of the processes submitted to the nodes. BatchJobs captures this information in a jobs subdirectory which is useful and pointed me to the source of the above error. Not sure whether this is available through BiocParallel? Again sorry for the unnecessary noise. Thomas On Tue, Sep 23, 2014 at 06:59:11PM -0700, Thomas Girke wrote:
Hi Valerie,
Thanks for looking into this.
Yes, if I include the bogus 'MYR' in *.tmpl then I am getting the same
error in R-release as well.
To double-check whether it is related to some nodes on our cluster (ours
has different node architectures and the IB interconnect can be flaky at
times), I restricted the computation to two specific nodes for all
comparisons using nodes="1:ppn=1+n02+n03". As you can see below, the same
computation works in R-release with both BiocParallel and BatchJobs. However,
if I run it in R-devel it only works with BatchJobs.
Certainly, there could still be another problem with our specfic
environment on the cluster, not sure?
For my specific application there is no rush to get things working in
BiocParallel right away. BatchJobs works fine for now.
Thomas
###############
## R-release ##
###############
library(BiocParallel); library(BatchJobs)
f <- function(i) system("hostname", intern=TRUE)
funs <- makeClusterFunctionsTorque("~/tmp/torque.tmpl")
param <- BatchJobsParam(4, resources=list(walltime="00:05:00", nodes="1:ppn=1+n02+n03", memory="1gb"), cluster.functions=funs)
register(param)
xx <- bplapply(1:4, f)
xx
xx
[[1]] [1] "n03" [[2]] [1] "n03" [[3]] [1] "n03" [[4]] [1] "n02" library(BatchJobs) loadConfig(conffile = "~/tmp/.BatchJobs.R") reg <- makeRegistry(id="BatchJobTest", work.dir="results") ids <- batchMap(reg, fun=f, 1:4) done <- submitJobs(reg, resources=list(walltime="00:05:00", nodes="1:ppn=1+n02+n03", memory="1gb")) sapply(1:4, function(x) loadResult(reg, x)) [1] "n03" "n03" "n03" "n02"
sessionInfo()
R version 3.1.0 (2014-04-10)
Platform: x86_64-unknown-linux-gnu (64-bit)
locale:
[1] C
attached base packages:
[1] stats graphics utils datasets grDevices methods base
other attached packages:
[1] BatchJobs_1.2 BBmisc_1.7 BiocParallel_0.6.1
loaded via a namespace (and not attached):
[1] BiocGenerics_0.10.0 DBI_0.2-7 RSQLite_0.11.4 Rcpp_0.11.2 brew_1.0-6 checkmate_1.0 codetools_0.2-8 digest_0.6.4 fail_1.2 foreach_1.4.2
[11] iterators_1.0.7 parallel_3.1.0 plyr_1.8.1 sendmailR_1.1-2 stringr_0.6.2 tools_3.1.0
#############
## R-devel ##
#############
library(BiocParallel); library(BatchJobs)
f <- function(i) system("hostname", intern=TRUE)
funs <- makeClusterFunctionsTorque("~/tmp/torque.tmpl")
param <- BatchJobsParam(4, resources=list(walltime="00:05:00", nodes="1:ppn=1+n02+n03", memory="1gb"), cluster.functions=funs)
register(param)
xx <- bplapply(1:4, f)
Error: 10 errors; first error:
For more information, use bplasterror(). To resume calculation, re-call the
function and set the argument 'BPRESUME' to TRUE or wrap the previous call in
bpresume().
bplasterror()
Error in vapply(head(which(is.error), n.print), f, character(1L)) :
values must be length 1, but FUN(X[[1]]) result is length 0
library(BatchJobs)
loadConfig(conffile = "~/tmp/.BatchJobs.R")
reg <- makeRegistry(id="BatchJobTest", work.dir="results")
ids <- batchMap(reg, fun=f, 1:4)
done <- submitJobs(reg, resources=list(walltime="00:05:00", nodes="1:ppn=1+n02+n03", memory="1gb"))
sapply(1:4, function(x) loadResult(reg, x))
[1] "n03" "n03" "n03" "n02"
sessionInfo()
R Under development (unstable) (2014-05-05 r65530)
Platform: x86_64-unknown-linux-gnu (64-bit)
locale:
[1] C
attached base packages:
[1] stats graphics utils datasets grDevices methods base
other attached packages:
[1] BatchJobs_1.3 BBmisc_1.7 BiocParallel_0.99.19
loaded via a namespace (and not attached):
[1] BiocGenerics_0.11.4 DBI_0.3.0 RSQLite_0.11.4 brew_1.0-6 checkmate_1.4 codetools_0.2-9 digest_0.6.4 fail_1.2
foreach_1.4.2 iterators_1.0.7
[11] parallel_3.2.0 sendmailR_1.1-2 stringr_0.6.2 tools_3.2.0
On Tue, Sep 23, 2014 at 09:41:44PM +0000, Valerie Obenchain wrote:
Hi, Martin and I looked into this a bit. It looks like a problem with handling an 'undefined error' returned from a worker (i.e., job did not run). When there is a problem executing the tmpl script no error message is sent back. The NULL is coerced to simpleError and becomes a problem downstream when the error processing is expecting messages of length > 0. You can reproduce the error by putting a typo in the script. For example replace R with something bogus such as MYR in this line: MYR CMD --no-save --no-restore "<%= rscript %>" /dev/stdout You said the script worked with release but not devel. Is it possible there's a problem with how R devel is being called on the cluster? Michel Lang (cc'd) implemented BatchJobs in BiocParallel. I'd like to get his opinion on how he wants to handle this type of error. Michel, let me know if you need more details, I can send another example off-line. Valerie On 09/22/2014 02:58 PM, Valerie Obenchain wrote:
Hi Thomas, Just wanted to let you know I saw this and am looking into it. Valerie On 09/20/2014 02:54 PM, Thomas Girke wrote:
Hi Martin, Micheal and Vincent, If I run the following code, with the release version of BiocParallel then it works (took me some time to actually realize that), but with the development version I am getting an error shown after the test code below. If I run the same test with BatchJobs from the devel branch alone then there is no problem. Thus, it seems there is some change in the devel version of BiocParallel causing this error? The torque.tmpl file I am using on our cluster is the standard one from BatchJobs here: https://github.com/tudo-r/BatchJobs/blob/master/examples/cfTorque/simple.tmpl For my application, I could stick with BatchJobs, but it would be nicer if I could get things to work with BiocParallel. Thanks, Thomas ############### ## Test Code ## ############### FUN <- function(i) system("hostname", intern=TRUE) library(BiocParallel); library(BatchJobs) funs <- makeClusterFunctionsTorque("torque.tmpl") param <- BatchJobsParam(4, resources=list(walltime="48:00:00", nodes="1:ppn=4", memory="4gb"), cluster.functions=funs) register(param) xx <- bplapply(1:4, FUN) Error: 4 errors; first error: For more information, use bplasterror(). To resume calculation, re-call the function and set the argument 'BPRESUME' to TRUE or wrap the previous call in bpresume()
bplasterror()
Error in vapply(head(which(is.error), n.print), f, character(1L)) :
values must be length 1,
but FUN(X[[1]]) result is length 0
sessionInfo()
R Under development (unstable) (2014-05-05 r65530) Platform: x86_64-unknown-linux-gnu (64-bit) locale: [1] C attached base packages: [1] stats graphics utils datasets grDevices methods base other attached packages: [1] BatchJobs_1.3 BBmisc_1.7 BiocParallel_0.99.19 loaded via a namespace (and not attached): [1] BiocGenerics_0.11.4 DBI_0.3.0 RSQLite_0.11.4 brew_1.0-6 checkmate_1.4 codetools_0.2-9 digest_0.6.4 fail_1.2 foreach_1.4.2 iterators_1.0.7 [11] parallel_3.2.0 sendmailR_1.1-2 stringr_0.6.2 tools_3.2.0
_______________________________________________ Bioc-devel at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/bioc-devel
_______________________________________________ Bioc-devel at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/bioc-devel