Choosing Intel MKL BLAS when compiling R

Interesting.  Incidentally, I mis-typed my link link below.  Usually I link to MKL using -lmkl_rt -liomp5 (without the -lpthread).  I'm still learning the vagaries of C++ and system libraries, but does this mean that I have been bypassing the Apple implementation?  If Intel's libiomp5 does not have this problem, would that explain why my code seems to be working?  Or is this an unrelated issue?

On Nov 2, 2011, at 8:39 AM, Simon Urbanek wrote:

On Nov 1, 2011, at 11:19 PM, Michael Braun wrote:

Simon:

Can you clarify your statement about OpenMP on OSX?  Is this an issue with OpenMP in general, or just R's use of it?  Or just Intel's use of it? 

It seems to be an issue in Apple's posix threads implementation that OpenMP relies on - it seems that the cost of thread synchronization is several orders of magnitude higher than on Linux which mean you can't use it for "short" tasks. It's deep enough the I don't think we can do anything about it. Luke Tierney has a good test case that shows the problem, I'll see if I can forward it here.

Cheers,
Simon

Michael


On Nov 1, 2011, at 9:09 PM, Simon Urbanek wrote:

On Nov 1, 2011, at 8:23 PM, Michael Braun wrote:

Jonathan:

Compiling R, and linking to Intel MKL, is quite simple, as long as you are using a recent version of MKL.  Intel has a MKL Link Advisor at

http://software.intel.com/en-us/articles/intel-mkl-link-line-advisor/

Choose the linking model you want, and then use the --with-blas option on configure.  For example, on my Mac, to link dynamically with the parallel version (which can be controlled with environment variables), I use the following:

'--with-blas=-L$(MKLROOT)/lib -lmkl_rt -lpthread -liomp5'

I compiled R 2.14.0 this morning with this option, and it seems to work fine.  I'm not sure what Dirk was referencing about how MKL is no longer compatible with R.  If there is an issue, it would be good to know about it.  Perhaps he can point us to the page in the documentation where this is mentioned.

The issue is not relevant on OS X because OpenMP performance on OS X for vector parallelization is very bad, so it is typically disabled.

One comment to the earlier discussion - it's not always as simple as replacing the BLAS back-end due to ABI difference in the different implementations. You'll notice, for example, when complex BLAS calls start crashing (because R can work around such issues at configure time but not if you switch BLAS). On Debian I'm usually compiling ATLAS by hand which yields the best performance so I don't run into that problem.

Cheers,
Simon

I hope this helps.

Michael


On Nov 1, 2011, at 7:59 PM, <r-sig-hpc-request at r-project.org> <r-sig-hpc-request at r-project.org> wrote:

Message: 1
Date: Tue, 1 Nov 2011 14:36:50 -0700
From: Jonathan Greenberg <jgrn at illinois.edu>
To: r-sig-hpc at r-project.org
Subject: [R-sig-hpc] Choosing Intel MKL BLAS when compiling R
Message-ID:
     <CABG0rfvnoswR+Hi7peaTAPczT_4j6FTV-H0rDLwqJOw05HrLuw at mail.gmail.com>
Content-Type: text/plain; charset=ISO-8859-1

r-sig-hpc'ers:

I am compiling R 2.14 from scratch on a cluster computer (I can't use
any package managers for the install, unfortunately), and they have
the Intel MKL libraries available for BLAS/LAPACK.  I have not
successfully gotten a configure/make/make install combo yet that uses
multiple CPUs to do, say:

a = matrix(rnorm(5000*5000), 5000, 5000)
b = matrix(rnorm(5000*5000), 5000, 5000)
# This following line, minimally, should be getting distributed.
c = a%*%b

First off, I've tried compiling it with the configure flags:
./configure --prefix=/myprefix/ --enable-BLAS-shlib

How do I then select the Intel MKL libraries to be used as the BLAS?
They are located in /usr/local/intel-11.1/mkl/lib/em64t

Thanks!

--j


--
Jonathan A. Greenberg, PhD
Assistant Professor
Department of Geography
University of Illinois at Urbana-Champaign
607 South Mathews Avenue, MC 150
Urbana, IL 61801
Phone: 415-763-5476
AIM: jgrn307, MSN: jgrn307 at hotmail.com, Gchat: jgrn307, Skype: jgrn3007
http://www.geog.illinois.edu/people/JonathanGreenberg.html

-------------------------------------------
Michael Braun
Associate Professor of Management Science (Marketing Group)
MIT Sloan School of Management
100 Main St.., E62-535
Cambridge, MA 02139
braunm at mit.edu
617-253-3436

-------------------------------------------
Michael Braun
Associate Professor of Management Science (Marketing Group)
MIT Sloan School of Management
100 Main St.., E62-535
Cambridge, MA 02139
braunm at mit.edu
617-253-3436

Michael Braun 
MIT Sloan School of Management
braunm at mit.edu
-------------------------------------------------- 
View my research at
http://braunm.scripts.mit.edu/