[Bioc-devel] Memory issues with BiocParallel::SnowParam()

Hi,

I ran my example code with SerialParam() which had a negligible 4%
memory increase between R 3.2.x and 3.1.x This 4% could very well
fluctuate a little bit and might be non significantly different from 0
if I run the test more times.

I also added a second example using code based on my analysis script.
With SerialParam(), the memory change is 13%, but with SnowParam()
it's 82% between the R versions mentioned already using 10 cores. It's
still far from the > 150% increase (2.5 fold change) I'm seeing with
the real data.

I initially thought that these observations ruled out everything else
except SnowParam(). However, maybe the initial 13% memory increase
multiplied by 10 (well, less then linear) is what I'm seeing with 10
cores (82% increase).

The updated information is available at
http://lcolladotor.github.io/SnowParam-memory/



As for what Vincent suggested of an AMI and EC2, I don't have
experience with them. I'm not sure I'll be able to look into them and
create a reproducible environment.


Cheers,
Leo

On Fri, Jul 10, 2015 at 7:12 AM, Vincent Carey
<stvjc at channing.harvard.edu> wrote:

I have had (potentially transient and environment-related) problems with
bplapply
in gQTLstats.   I substituted the foreach abstractions and the code worked.
I still
have difficulty seeing how to diagnose the trouble I ran into.

I'd suggest that you code so that you can easily substitute parallel- or
foreach- or
BatchJobs-based cluster control.  This can help crudely isolate the source
of trouble.

It would be very nice to have a way of measuring resource usage in cluster
settings,
both for diagnosis and strategy selection.  For jobs that succeed, BatchJobs
records
memory used in its registry database, based on gc().  I would hope that
there are
tools that could be used to help one figure out how to factor a task so that
it is feasible
given some view of environment constraints.

It might be useful for you to build an AMI and then a cluster that allows
replication of
the condition you are seeing on EC2.  This could help with diagnosis and
might be
a basis for defining better instrumentation tools for both diagnosis and
planning.

On Fri, Jul 10, 2015 at 12:23 AM, Leonardo Collado Torres <lcollado at jhu.edu>
wrote:

Hi,

I have a script that at some point generates a list of DataFrame
objects which are rather large matrices. I then feed this list to
BiocParallel::bplapply() and process them.

Previously, I noticed that in our SGE managed cluster using
MulticoreParam() lead to 5 to 8 times higher memory usage as I posted
in https://support.bioconductor.org/p/62551/#62877. Martin posted in
https://support.bioconductor.org/p/62551/#62880 that "Probably the
tools used to assess memory usage are misleading you." This could be
true, but they are the tools that determine memory usage for all jobs
in the cluster. Meaning that if my memory usage blows up according to
these tools, my jobs get killed.

That was with R 3.1.x and in particular running

https://github.com/leekgroup/derSoftware/blob/gh-pages/step1-fullCoverage.sh
with

$ sh step1-fullCoverage.sh brainspan

which at the time (Nov 4th, 2014) used 173.5 GB of RAM with 10 cores.
I recently tried to reproduce this (to check changes in run time given
rtracklayer's improvements with BigWig files) using R 3.2.x and the
memory went up to 450 GB before the job got killed given the maximum
memory I specified for the job. The same is true using R 3.2.0.

Between R 3.1.x and 3.2.0, `derfinder` is nearly identical (just one
bug fix is different, for other code not used in this script). I know
that BiocParallel changed quite a bit between those versions, and in
particular SnowParam(). So that's why my prime suspect is
BiocParallel.

I made a smaller reproducible example which you can view at
http://lcolladotor.github.io/SnowParam-memory/. This example uses a
list of data frames with random data, and also uses 10 cores. You can
see there that in R versions 3.1.x, 3.2.0 and 3.2.x, MulticoreParam()
does use more memory than SnowParam(), as reported by SGE. Beyond the
actual session info differences due to changes in BiocParalell's
implementation, I noticed that the cluster type changed from PSOCK to
SOCK. I ignore if this could explain the memory increase.

The example doesn't generate the huge fold change between R 3.1.x and
the other two versions (still 1.27x > 1x) that I see with my analysis
script, so in that sense it's not the best example for the problem I'm
observing. My tests with

https://github.com/leekgroup/derSoftware/blob/gh-pages/step1-fullCoverage.sh
were between June 23rd and 28th, so maybe some recent changes in
BiocParallel addressed this issue.


I'm not sure how to proceed now. One idea is to make another example
with the same type of objects and operations I use in my analysis
script.

A second one is to run my analysis script with SerialParam() on the
different R versions to check if they use different amounts of memory
which would suggest that the memory issue is not caused by
SnowParam(). For example, maybe changes in rtracklayer are the ones
driving the huge memory changes I'm seeing in my analysis scripts.

However, I don't really suspect rtracklayer given the memory load
reported that I checked manually a couple of times with "qmem". I
believe that the memory blows up at

https://github.com/leekgroup/derSoftware/blob/gh-pages/step1-fullCoverage.R#L124
which in turn uses derfinder::filterData(). This function imports:

'[', '[<-', '[[', colnames, 'colnames<-', lapply methods from IRanges
Rle, DataFrame from S4Vectors
Reduce method from S4Vectors


https://github.com/lcolladotor/derfinder/blob/master/R/filterData.R#L49-L51


Best,
Leo


History of analysis scripts doesn't reveal any other leads

https://github.com/leekgroup/derSoftware/commits/gh-pages/step1-fullCoverage.sh

https://github.com/leekgroup/derSoftware/commits/gh-pages/step1-fullCoverage.R