technical:whitepaper:r-runtime-blas-lapack

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision		 Previous revision
technical:whitepaper:r-runtime-blas-lapack [2018-12-10 11:37] – [Substituting an alternate library] freytechnical:whitepaper:r-runtime-blas-lapack [2018-12-10 12:40] (current) – [R: Runtime-configuration BLAS/LAPACK] frey
Line 1: Line 1:
 +====== R: Runtime-configuration BLAS/LAPACK ======
  
 +The R Project for Statistical Computing is used on our clusters by a wide variety of scientific disciplines.  Though the breadth of applications is wide, many of them require the functionality of BLAS/LAPACK libraries.  R provides its own baseline implementations that will build on any system; naturally, one cannot expect these BLAS/LAPACK libraries to be highly performant relative to implementations like:
 +  * Intel Math Kernel Library (MKL)
 +  * Automatically-Tuned Linear Algebra Software (ATLAS)
 +The build procedure for R allows the package to be configured for building against external BLAS/LAPACK libraries.  Once the base R build has completed and the resulting software has been installed, additional R libraries can be configured and installed atop it.  It has been noted in the past that:
 +  - Producing //N// such builds of R that vary only in the choice of underlying BLAS/LAPACK:
 +    * can require on the order of //N// times the disk space of a single build
 +    * puts a greater burden on the sysadmin to maintain all //N// similarly-outfitted copies
 +  - R only makes use of standardized BLAS/LAPACK APIs, so any standard BLAS/LAPACK library should be able to be chosen at runtime (not just build time).
 +
 +===== Substituting an alternate library =====
 +
 +Others have published articles in the past detailing the substitution of the ATLAS library by doing the following to a basic R build (which was built with its bundled BLAS/LAPACK):
 +  * [[https://www.r-bloggers.com/r-r-with-atlas-r-with-openblas-and-revolution-r-open-which-is-fastest/]]
 +  * [[https://stackoverflow.com/questions/29984141/does-installing-blas-atlas-mkl-openblas-will-speed-up-r-package-that-is-written]]
 +The basic idea is:
 +  * copy ''libatlas.so'' to ''** R_HOME**/lib64/R/lib''
 +  * remove ''libRblas.so'' and ''libRlapack.so'' from ''** R_HOME**/lib64/R/lib''
 +  * symlink ''libRblas.so'' and ''libRlapack.so'' to ''libatlas.so'' in ''** R_HOME**/lib64/R/lib''
 +This copy of R is configured to use ''** R_HOME**/lib64/R/lib'' to resolve shared libraries, so when executing the ''R'' command, for example, the symlinks will lead the runtime linker to the ATLAS library when resolving BLAS/LAPACK functions.
 +
 +This scheme requires two things:
 +  - the user must have ownership of the R installation or sufficient privileges to alter the files
 +  - the BLAS/LAPACK substitution will happen on time only (probably shortly after the library is built)
 +While the first condition is obvious, the second may not seem important, especially for a build of R being maintained by an arbitrary user in an arbitrary location on the filesystem.  However, computational reproducibility would demand that any alteration to the underlying BLAS/LAPACK be present -- or at least able to be restored -- at any time.  This is one reason why ''libatlas.so'' was copied into the build and symlinks were used:  having other BLAS/LAPACK libraries present, the ''libRblas.so'' and ''libRlapack.so'' symlinks can be altered as necessary.  The caveat, however, is that:
 +  * only a single choice of underlying BLAS/LAPACK can be active
 +  * the underlying BLAS/LAPACK can be changed only when that build of R is not being executed/used
 +
 +A simple way to organize multiple underlying BLAS/LAPACK libraries in a single R installation is to create subdirectories for each variant:
 +
 +^Path  ^Description^
 +|''** R_HOME**/lib64/R/lib''  |base directory where R looks for shared libraries by default|
 +|''** R_HOME**/lib64/R/lib/libRblas.so''  |symlink to chosen BLAS library (from one of the subdirectories herein)|
 +|''** R_HOME**/lib64/R/lib/libRlapack.so''  |symlink to chosen LAPACK library (from one of the subdirectories herein)|
 +|''** R_HOME**/lib64/R/lib/atlas''  |directory to hold ''libatlas.so''|
 +|''** R_HOME**/lib64/R/lib/rblas''  |directory to hold the bundled ''libRblas.so'' and ''libRlapack.so'' produced by R build procedure|
 +|''** R_HOME**/lib64/R/lib/mkl''  |directory to hold MKL variants|
 +|''** R_HOME**/lib64/R/lib/mkl/seq''  |directory to hold sequential MKL variant|
 +|''** R_HOME**/lib64/R/lib/mkl/thr''  |directory to hold threaded MKL variant|
 +
 +==== R BLAS/LAPACK ====
 +
 +When we restructured the R ''lib64/R/lib'' directory, the bundled ''libRblas.so'' and ''libRlapack.so'' shared library files were moved to the ''rblas'' subdirectory.  To configure R to use its bundled libraries:
 +
 +<code bash>
 +$ cd ${R_HOME}/lib64/R/lib
 +$ rm -f libR{blas,lapack}.so
 +$ ln -s rblas/libRblas.so .
 +$ ln -s rblas/libRlapack.so .
 +</code>
 +
 +==== ATLAS ====
 +
 +The ATLAS library contains both BLAS and LAPACK APIs in a single shared library.  With ''libatlas.so'' copied into the ''** R_HOME**/lib64/R/lib/atlas'' subdirectory, we configure R to use ATLAS:
 +
 +<code bash>
 +$ cd ${R_HOME}/lib64/R/lib
 +$ rm -f libR{blas,lapack}.so
 +$ ln -s atlas/libRblas.so .
 +$ ln -s atlas/libRlapack.so .
 +</code>
 +
 +==== Sequential MKL ====
 +
 +A C source file containing a dummy function was created in ''** R_HOME**/lib64/R/lib/mkl/seq'':
 +
 +<file C shim.c>
 +int
 +mkl_shim_dummy(void)
 +{
 + return 0;
 +}
 +</file>
 +
 +The shim library is then created thusly:
 +
 +<code bash>
 +$ cd ${R_HOME}/lib64/R/lib/mkl/seq
 +$ icc -shared -o libRblas.so -mkl=sequential shim.c
 +$ ln -s libRblas.so libRlapack.so
 +</code>
 +
 +To configure R to use the sequential MKL:
 +
 +<code bash>
 +$ cd ${R_HOME}/lib64/R/lib
 +$ rm -f libR{blas,lapack}.so
 +$ ln -s mkl/seq/libRblas.so .
 +$ ln -s mkl/seq/libRlapack.so .
 +</code>
 +
 +==== Threaded MKL ====
 +
 +A C source file containing a dummy function was created in ''** R_HOME**/lib64/R/lib/mkl/thr'':
 +
 +<file C shim.c>
 +int
 +mkl_shim_dummy(void)
 +{
 + return 0;
 +}
 +</file>
 +
 +Since our R build used the GNU C compiler, the threaded MKL variant only works if the shim library is built against the GNU OpenMP runtime.  Using just "-mkl=parallel" links against the Intel OpenMP runtime which in testing yielded numerical issues (not actual crashes).  The shim library is then created thusly:
 +
 +<code bash>
 +$ cd ${R_HOME}/lib64/R/lib/mkl/thr
 +$ icc -shared -o libRblas.so shim.c -lmkl_gnu_thread -lmkl_core -lmkl_intel_lp64
 +$ ln -s libRblas.so libRlapack.so
 +</code>
 +
 +To configure R to use the threaded MKL:
 +
 +<code bash>
 +$ cd ${R_HOME}/lib64/R/lib
 +$ rm -f libR{blas,lapack}.so
 +$ ln -s mkl/thr/libRblas.so .
 +$ ln -s mkl/thr/libRlapack.so .
 +</code>
 +
 +===== Runtime-configurable substitution =====
 +
 +By stashing each BLAS/LAPACK variant in its own subdirectory, our copy of R is actually fairly close to being runtime-configurable with respect to choice of BLAS/LAPACK.  Since all R commands will setup the environment to have the runtime linker check ''** R_HOME**/lib64/R/lib'' for shared libraries, the ''libRblas.so'' and ''libRlapack.so'' symlinks in that directory will always have priority over any other path we might add to ''LD_LIBRARY_PATH'' prior to issuing the ''R'' command, for example.  However, if ''libRblas.so'' and ''libRlapack.so'' are **not** present in that directory, the runtime linker will be forced to consult other paths present in ''LD_LIBRARY_PATH''.
 +
 +On our Caviness cluster we include no BLAS/LAPACK library symlinks in the base directory which R checks for shared libraries:
 +
 +<code bash>
 +$ cd ${R_HOME}/lib64/R/lib
 +$ rm -f libR{blas,lapack}.so
 +</code>
 +
 +In our VALET package definition for R we configure four variants that differ by BLAS/LAPACK (as a feature tag)
 +
 +<code bash>
 +$ vpkg_versions r
 +   :
 +r                The R Project for Statistical Computing
 +  3.5            alias to r/3.5.1
 +* 3.5.1          R 3.5.1 with system compilers, ATLAS
 +  3.5.1:mkl-seq  R 3.5.1 with system compilers, MKL (sequential)
 +  3.5.1:mkl-thr  R 3.5.1 with system compilers, MKL (multithread)
 +  3.5.1:rblas    R 3.5.1 with system compilers, R reference BLAS/LAPACK
 +</code>
 +
 +with ATLAS as the default (recommended) choice of underlying BLAS/LAPACK.  Each variant of the 3.5.1 version uses the same installation prefix, but adds a unique BLAS/LAPACK subdirectory to ''LD_LIBRARY_PATH'' when added to the user's environment:
 +
 +<code bash>
 +$ vpkg_require r/3.5.1
 +Adding package `r/3.5.1` to your environment
 +$ echo $LD_LIBRARY_PATH
 +/opt/shared/r/3.5.1/lib64/R/lib/atlas: ...
 +$ vpkg_rollback all
 +$ vpkg_require r/3.5.1:rblas
 +Adding package `r/3.5.1:rblas` to your environment
 +$ echo $LD_LIBRARY_PATH
 +/opt/shared/r/3.5.1/lib64/R/lib/rblas:
 +</code>
 +
 +This works fine so long as you don't attempt to install any R modules that require the BLAS/LAPACK functionality.  The R module-building environment defaults to using link search paths in the base directory (''** R_HOME**/lib64/R/lib'') and fails to find ''libRblas.so'' or ''libRlapack.so''.  This is easily fixed by altering two lines in the R installation's standard make flags file at ''** R_HOME**/lib64/R/etc/Makeconf'':
 +
 +<code>
 +BLAS_LIBS = -L"$(R_HOME)/lib$(R_ARCH)/$(R_BLAS_VARIANT)" -lRblas
 +</code>
 +
 +and
 +
 +<code>
 +LAPACK_LIBS = -L"$(R_HOME)/lib$(R_ARCH)/$(R_BLAS_VARIANT)" -lRlapack
 +</code>
 +
 +The link search added here will default to the usual path if ''R_BLAS_VARIANT'' is not set in the user's environment.  But in our VALET configuration of each of the variants, we set the appropriate relative path for ''R_BLAS_VARIANT'':
 +
 +<code bash>
 +$ vpkg_require r/3.5.1:rblas
 +Adding package `r/3.5.1:rblas` to your environment
 +$ echo $R_BLAS_VARIANT 
 +rblas
 +$ vpkg_rollback all
 +$ vpkg_require r/3.5.1:mkl-seq
 +Adding package `r/3.5.1:mkl-seq` to your environment
 +$ echo $R_BLAS_VARIANT 
 +mkl/seq
 +</code>
 +
 +We now have a runtime-configurable BLAS/LAPACK for this single installation of R, and any properly-packaged R modules should build fine against it.
 +
 +===== VALET configuration =====
 +
 +Here is the VALET configuration for our installation of R 3.5.1 with runtime-configurable BLAS/LAPACK:
 +
 +<file yaml>
 +r:
 +  description:       The R Project for Statistical Computing
 +  url:               http://www.r-project.org/
 +  prefix:            /opt/shared/r
 +
 +  default-version:   "3.5.1"
 +
 +  versions:
 +    "3.5":
 +      alias-to:      3.5.1
 +    "3.5.1":
 +      description:   R 3.5.1 with system compilers, ATLAS
 +      actions:
 +        - libdir:    lib64/R/lib
 +        - incdir:    lib64/R/include
 +        - libdir:    lib64/R/lib/atlas
 +        - variable:  R_BLAS_VARIANT
 +          operator:  set
 +          value:     atlas
 +
 +    "3.5.1:rblas":
 +      description:   R 3.5.1 with system compilers, R reference BLAS/LAPACK
 +      prefix:        3.5.1
 +      actions:
 +        - libdir:    lib64/R/lib
 +        - incdir:    lib64/R/include
 +        - libdir:    lib64/R/lib/rblas
 +        - variable:  R_BLAS_VARIANT
 +          operator:  set
 +          value:     rblas
 +
 +    "3.5.1:mkl-seq":
 +      description:   R 3.5.1 with system compilers, MKL (sequential)
 +      prefix:        3.5.1
 +      actions:
 +        - libdir:    lib64/R/lib
 +        - incdir:    lib64/R/include
 +        - libdir:    lib64/R/lib/mkl/seq
 +        - variable:  R_BLAS_VARIANT
 +          operator:  set
 +          value:     mkl/seq
 +
 +    "3.5.1:mkl-thr":
 +      description:   R 3.5.1 with system compilers, MKL (multithread)
 +      prefix:        3.5.1
 +      actions:
 +        - libdir:    lib64/R/lib
 +        - incdir:    lib64/R/include
 +        - libdir:    lib64/R/lib/mkl/thr
 +        - variable:  R_BLAS_VARIANT
 +          operator:  set
 +          value:     mkl/thr
 +</file>