Skip to content

Rsge combined with mclapply?

5 messages · Daniel Sabanés Bové, Kasper Daniel Hansen

Original
Daniel Sabanés Bové
# 
Dear Dan Bode,

first of all I would like to thank you very much for your R-package 
"Rsge". It looks very helpful, because I need to submit my parallel jobs 
on the local cluster using SGE.

Actually I have a trivially parallel problem which could be sequentially 
solved by lapply(), and on a single machine with multiple cores by 
mclapply() from the base R-package "parallel". Because the local cluster 
has nodes with multiple cores, I would like to combine sge.parLapply (to 
distribute to nodes) with mclapply (not lapply, to distribute to cores 
in each node).

Before doing some modifications to sge.parParApply like introducing 
another apply.method for mclapply, I would like to be sure that you have 
not already programmed a nice solution for this problem. So, did you 
already have the same problem as I am facing now? Would you think 
another apply.method for sge.parParApply would be a good way to go?

Thank you very much in advance for any hints or suggestions.

Best regards
Daniel Sabanes
8 days later
Kasper Daniel Hansen
# 
I am not too familiar with Rsge.  But on our SGE cluster, the sysadmin
has a specific type of parallel environment where you can do
  qsub -pe local 8-12
which asks for 8-12 cores on the same node.  My guess is that once
such a parallel resource has been set up, it should be trivial to use
with Rsge.

Kasper

On Tue, Jun 5, 2012 at 6:19 AM, Daniel Saban?s Bov?
<daniel.sabanesbove at ifspm.uzh.ch> wrote:
Daniel Sabanés Bové
# 
Hi Kasper,

thank you for this hint! I was not aware of this possibility. Yes, 
indeed, then it should be trivial to let multiple jobs run on cores of 
the same node. In the meantime I have found a solution by tweaking the 
Rsge package to use mclapply() from the base R-package "parallel". This 
also works.

Best regards
Daniel


-------- Original-Nachricht --------
Betreff: Re: [R-sig-hpc] Rsge combined with mclapply?
Von: Kasper Daniel Hansen <kasperdanielhansen at gmail.com>
An: Daniel Saban?s Bov? <daniel.sabanesbove at ifspm.uzh.ch>
Datum: Mit 13 Jun 2012 23:10:23 CEST
Kasper Daniel Hansen
# 
On Thu, Jun 14, 2012 at 4:42 AM, Daniel Saban?s Bov?
<daniel.sabanesbove at ifspm.uzh.ch> wrote:
This might work in the sense that it produces workable code on your
system.  But are you sure that you do not over-use your resources.
For example, on our system, I can qsub a job that SGE thinks takes up
a single slot (= core, the default) and then still use mclapply over 8
cores.  In that situation, I have asked for just 1 core, but I am
really using 8.  This has clear implications for the other users on
the system as well as resource allocation.  This is because our SGE
system is not too tightly locked down - it may be possible for a
sysadmin to prevent this, but I know I can do it on the system I have
access to.

Kasper
Daniel Sabanés Bové
# 
Hi Kasper,

OK, I understand. If the cluster is organised like that, you do not have 
to tinker with mclapply anyway.

But on our cluster the whole node is allocated for one job, so I do not 
have this problem. Or, put differently, if I just used the standard Rsge 
function sge.parLapply, I would have the problem that 8 cores are 
allocated for one single R job. This was my motivation for writing an 
sge.parMClapply function.

Thanks for the clarification!

Best
Daniel

-------- Original-Nachricht --------
Betreff: Re: [R-sig-hpc] Rsge combined with mclapply?
Von: Kasper Daniel Hansen <kasperdanielhansen at gmail.com>
An: Daniel Saban?s Bov? <daniel.sabanesbove at ifspm.uzh.ch>
Datum: Don 14 Jun 2012 14:27:13 CEST