On 09/21/2016 10:42 AM, Christian Arnold wrote:
Dear Bioconductor developers,
I am having a somewhat mysterious and challenging problem which we
believe is a bug in either (1) BiocParallel or (2) the lpsymphony
library from either Bioconductor or the SYMPHONY backend.
Thank you for the clear report
I think the problem is that lpsymphony is compiled by default to use
openmp, and this parallelization environment interferes with the fork
processes that BiocParallel::MulticoreParam() uses.
I was lead to this conclusion by using the 'gdb' debugger and
interrupting R when it hung -- there were a number of openmp threads
persisting, as in the abbreviated session below
$ Rdev -d gdb
GNU gdb (Ubuntu 7.11.1-0ubuntu1~16.04) 7.11.1
Copyright (C) 2016 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later
<http://gnu.org/licenses/gpl.html>
...
(gdb) r
...
> source("lp.R", echo=T, max=Inf) ## your script
> library(BiocParallel)
> library(lpsymphony)
> lpsymphonyTest <- function(x) {
+ ## Example from the R help of lpsymphony_solve_LP
+ obj <- c(2, 4, 3)
+ mat <- matrix(c(3, 2, 1, 4, 1, 3, 2, 2, 2), nrow = 3)
+ dir <- c("<=", "<=", "<=")
+ rhs <- c(60, 40, 80)
+ max <- TRUE
+ test = lpsymphony_solve_LP(obj, mat, dir, rhs, max = max)
+ Sys.sleep(1)
+ return(1)
+ }
> nIterationsAndCores = 2
> register(MulticoreParam(workers = nIterationsAndCores))
> ## register(bpstart(MulticoreParam(workers = nIterationsAndCores)))
>
> # First run: Two iterations, should run in parallel
> commonGenes <- bplapply(X = seq(2), FUN = lpsymphonyTest)
> # Second run: One iteration, should run on only one core
> commonGenes <- bplapply(X = seq(1), FUN = lpsymphonyTest)
[New Thread 0x7fffeb483700 (LWP 29894)]
[New Thread 0x7fffeac82700 (LWP 29895)]
[New Thread 0x7fffea481700 (LWP 29896)]
[New Thread 0x7fffe9c80700 (LWP 29897)]
[New Thread 0x7fffe947f700 (LWP 29898)]
[New Thread 0x7fffe8c7e700 (LWP 29899)]
[New Thread 0x7fffe847d700 (LWP 29900)]
> # Third run: Two iterations again, should run in parallel but crashed
and never finishes
> commonGenes <- bplapply(X = seq(2), FUN = lpsymphonyTest)
^C
...
(gdb) info threads
Id Target Id Frame
* 1 Thread 0x7ffff7fc97c0 (LWP 29865) "R" 0x00007ffff5a91d13 in
select () at ../sysdeps/unix/syscall-template.S:84
2 Thread 0x7fffeb483700 (LWP 29894) "R" 0x00007ffff5f8cb4f in ??
() from /usr/lib/x86_64-linux-gnu/libgomp.so.1
3 Thread 0x7fffeac82700 (LWP 29895) "R" 0x00007ffff5f8cb4f in ??
() from /usr/lib/x86_64-linux-gnu/libgomp.so.1
4 Thread 0x7fffea481700 (LWP 29896) "R" 0x00007ffff5f8cb4f in ??
() from /usr/lib/x86_64-linux-gnu/libgomp.so.1
5 Thread 0x7fffe9c80700 (LWP 29897) "R" 0x00007ffff5f8cb4f in ??
() from /usr/lib/x86_64-linux-gnu/libgomp.so.1
6 Thread 0x7fffe947f700 (LWP 29898) "R" 0x00007ffff5f8cb4f in ??
() from /usr/lib/x86_64-linux-gnu/libgomp.so.1
7 Thread 0x7fffe8c7e700 (LWP 29899) "R" 0x00007ffff5f8cb4f in ??
() from /usr/lib/x86_64-linux-gnu/libgomp.so.1
8 Thread 0x7fffe847d700 (LWP 29900) "R" 0x00007ffff5f8cb4f in ??
() from /usr/lib/x86_64-linux-gnu/libgomp.so.1
I compiled lpsymphony without openmp by changing the top-level
lpsymphony/configure file to explicitly exclude openmp
$ svn diff lpsymphony/configure
Index: lpsymphony/configure
===================================================================
--- lpsymphony/configure (revision 121222)
+++ lpsymphony/configure (working copy)
@@ -123,6 +123,7 @@
else
(cd src/SYMPHONY && \
./configure \
+ --enable-openmp=no \
--enable-static --disable-shared --with-pic \
--with-application=no --disable-dependency-tracking \
--disable-zlib --disable-bzlib \
and then installing with R CMD INSTALL lpsymphony. Your test script then
succeeded; providing weak verification that openmp was the problem.
There are several things.
1. since SYMPHONY is already using openmp, and it is a sophisticated
piece of software, probably there is little value to using BiocParallel
on top of it?
2. It seems that one can use SnowParam() effectively; this requires
modifying lpsymphonyTest() to require(lpsymphony)
lpsymphonyTest <- function(x) {
## Example from the R help of lpsymphony_solve_LP
require(lpsymphony)
...
The cost of starting the independent processes can be amortized across a
session by starting them manually at the beginning (and stopping them at
the end)
param = bpstart(SnowParam(workers = nIterationsAndCores))
register(param)
...
bpstop(param)
3. It would be good if the maintainer of lpsymphony exposed the
enable-openmp (and other?) flags so that one could R CMD INSTALL
--configure-args="--enable-openmp=no" lpsymphony
4. I'll work to come up with a simpler example and see if I can make
BiocParallel robust to use of openmp; this might be challenging for me.
Hope that helps, if does not solve, the problem, and thank you for the
report.
Martin
The problem is, in a nutshell, that R silently crashes or, to be more
precise, never finishes the calculation when using the only function
from lpsymphony in combination with BiocParallel.
We updated the latest version of lpsymphony last week, did not help.
This can be reproduced with two different configurations, using
different library versions of both BiocParallel and lpsymphony,
respectively. We also tested this on two different machines (one of
which runs R in the devel version, one does not; see the two
sessionInfo() at the end of this message)
The following code can reproduce the problem:
/library(BiocParallel)/
/library(lpsymphony)/
/lpsymphonyTest <- function(x){/
/# Example from the R help of lpsymphony_solve_LP/
/obj <- c(2, 4, 3)/
/mat <- matrix(c(3, 2, 1, 4, 1, 3, 2, 2, 2), nrow = 3)/
/dir <- c("<=", "<=", "<=")/
/rhs <- c(60, 40, 80)/
/max <- TRUE/
/test = lpsymphony_solve_LP(obj, mat, dir, rhs, max = max)/
/Sys.sleep(1)/
/return(1)/
/}/
/# First run: Two iterations, should run in parallel/
/nIterationsAndCores = 2/
/register(MulticoreParam(workers = nIterationsAndCores))/
/commonGenes <- bplapply(X = seq(2), FUN = lpsymphonyTest) /
/# Second run: One iteration, should run on only one core/
/commonGenes <- bplapply(X = seq(1), FUN = lpsymphonyTest) /
/# Third run: Two iterations again, should run in parallel but crashed
and never finishes/
/commonGenes <- bplapply(X = seq(2), FUN = lpsymphonyTest) /
What happens in our case is that the third run never finishes. The two R
processes are in state “SLEEP” rather than running. Once you execute the
bplapply loop for only one iteration, everything after with number of
iterations > 1 crashes. It works fine for any number of iterations > 1
despite of the order, but again, executing only a single iteration
causes the problems afterwards. Note that using registering again for
the second and third run does not help and yields the same behavior.
In fact, we have another issue, namely that the parallelization with
BiocParallel in combination with the IHW package (which calls lpsymphony
in the background) does not yield running times as expected but instead
much, much longer ones. However, let's focus on this one first, because
we think they might be related.
Thanks for your help,
Best
Christian
These are the two configurations where this problem can be reproduced:
(1)
/R Under development (unstable) (2016-06-30 r70858)/
//
/Platform: x86_64-pc-linux-gnu (64-bit)/
//
/Running under: Ubuntu 16.04.1 LTS/
//
/
/
//
/locale:/
//
/[1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C LC_TIME=de_DE.UTF-8
LC_COLLATE=en_US.UTF-8 /
//
/[5] LC_MONETARY=de_DE.UTF-8 LC_MESSAGES=en_US.UTF-8
LC_PAPER=de_DE.UTF-8 LC_NAME=C /
//
/[9] LC_ADDRESS=C LC_TELEPHONE=C LC_MEASUREMENT=de_DE.UTF-8
LC_IDENTIFICATION=C /
//
/attached base packages:/
//
/[1] stats graphics grDevices utils datasets methods base /
//
/other attached packages:/
//
/[1] lpsymphony_1.1.2 BiocParallel_1.7.8/
//
/loaded via a namespace (and not attached):/
//
/[1] parallel_3.4.0 tools_3.4.0 /
(2)
/R version 3.3.1 (2016-06-21)/
//
/Platform: x86_64-pc-linux-gnu (64-bit)/
//
/Running under: CentOS release 6.5 (Final)/
//
/
//locale:/
//
/[1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C LC_TIME=en_US.UTF-8 /
//
/[4] LC_COLLATE=en_US.UTF-8 LC_MONETARY=en_US.UTF-8
LC_MESSAGES=en_US.UTF-8 /
//
/[7] LC_PAPER=en_US.UTF-8 LC_NAME=C LC_ADDRESS=C /
//
/[10] LC_TELEPHONE=C LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C /
//
/
//attached base packages:/
//
/[1] stats graphics grDevices utils datasets methods base /
//
/
//other attached packages:/
//
/[1] lpsymphony_1.0.2 BiocParallel_1.6.6/
//
/
//loaded via a namespace (and not attached):/
//
/[1] parallel_3.3.1 tools_3.3.1 /
[[alternative HTML version deleted]]
_______________________________________________
Bioc-devel@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/bioc-devel
This email message may contain legally privileged and/or...{{dropped:2}}
_______________________________________________
Bioc-devel@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/bioc-devel
