R devel - Aug 2017 - Control multi-threading in standard matrix product

Dear R Core Team,

I wish to report what can be viewed as a bug or at least a strange 
behavior in R-3.4.1. I ask my question here (as recommended on 
https://www.r-project.org/bugs.html) since I am not member of the R's 
Bugzilla.

When running 'R --vanilla' from the command line, the standard matrix 
product is by default based on BLAS and multi-threaded on all cores 
available on the machine, c.f. following examples:

n=10000
p=1000
q=5000
A = matrix(runif(n*p),nrow=n, ncol=p)
B = matrix(runif(p*q),nrow=p, ncol=q)
C = A %*% B # multi-threaded matrix product


However, the default behavior to use all available cores can be an 
issue, especially on shared computing resources or when the matrix 
product is used in parallelized section of codes (for instance with 
'mclapply' from the 'parallel' package). For instance, the
default
matrix product is single-threaded in R-3.3.2 (I ran a test on my 
machine), this new features will deeply affect the behavior of existing 
R packages that use other multi-threading solutions.

Thanks to this stackoverflow question 
(https://stackoverflow.com/questions/45794290/in-r-how-to-control-multi-threading-in-blas-parallel-matrix-product),
I now know that it is possible to control the number of BLAS threads 
thanks to the package 'RhpcBLASctl'. However, being able to control the 
number of threads should maybe not require to use an additional package.

In addition, the doc 'matmult' does not mention this point, it points to
the 'options' doc page and especially the 'matprod' section, in
which
the multi-threading is not discussed.


Here is the results of the 'sessionInfo()' function on my machine for 
R-3.4.1:
R version 3.4.1 (2017-06-30)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Ubuntu 16.04.3 LTS

Matrix products: default
BLAS: /usr/lib/openblas-base/libblas.so.3
LAPACK: /usr/lib/libopenblasp-r0.2.18.so

locale:
  [1] LC_CTYPE=fr_FR.utf8       LC_NUMERIC=C
  [3] LC_TIME=fr_FR.utf8        LC_COLLATE=fr_FR.utf8
  [5] LC_MONETARY=fr_FR.utf8    LC_MESSAGES=fr_FR.utf8
  [7] LC_PAPER=fr_FR.utf8       LC_NAME=C
  [9] LC_ADDRESS=C              LC_TELEPHONE=C
[11] LC_MEASUREMENT=fr_FR.utf8 LC_IDENTIFICATION=C

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods base

loaded via a namespace (and not attached):
[1] compiler_3.4.1



and for R-3.3.2:
R version 3.3.2 (2016-10-31)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Ubuntu 16.04.3 LTS

locale:
  [1] LC_CTYPE=fr_FR.utf8       LC_NUMERIC=C
  [3] LC_TIME=fr_FR.utf8        LC_COLLATE=fr_FR.utf8
  [5] LC_MONETARY=fr_FR.utf8    LC_MESSAGES=fr_FR.utf8
  [7] LC_PAPER=fr_FR.utf8       LC_NAME=C
  [9] LC_ADDRESS=C              LC_TELEPHONE=C
[11] LC_MEASUREMENT=fr_FR.utf8 LC_IDENTIFICATION=C

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods base


Thanks in advance,
Best regards
||

-- 
Ghislain Durif
--------------------------
Research engineer THOTH TEAM
INRIA Grenoble Alpes (France)


	[[alternative HTML version deleted]]

Hi Ghislain,

I think you might be comparing two versions of R with different BLAS 
implementations, one that is single threaded (is your 3.3.2 used with 
reference blas?) and one that is multi threaded (3.4.1 with openblas). 
Could you check with "perf"? E.g. run your benchmark with "perf
record"
in both cases and you should see the names of the hot BLAS functions and 
this should reveal the BLAS implementation (look for dgemm).

In Ubuntu, if you install R from the package system, whenever you run it 
it will use the BLAS currently installed via the package system. However 
if you build R from source on Ubuntu, by default, it will use the 
reference BLAS which is distributed with R. Section "Linear algebra"
of
"R Installation and Administration" has details on how to build R with
different BLAS/LAPACK implementations.

Sadly there is no standard way to specify the number of BLAS worker 
threads. RhpcBLASctl has specific code for several existing 
implementations, but R itself does not attempt to control BLAS multi 
threading in any way. It is expected the user/system administrator will 
configure their BLAS implementation of choice to use the number of 
threads they need. A similar problem exists in other internally 
multi-threaded third-party libraries, used by packages - R cannot 
control how many threads they run.

Best
Tomas

On 08/21/2017 02:55 PM, Ghislain Durif wrote:
> Dear R Core Team,
>
> I wish to report what can be viewed as a bug or at least a strange
> behavior in R-3.4.1. I ask my question here (as recommended on
> https://www.r-project.org/bugs.html) since I am not member of the R's
> Bugzilla.
>
> When running 'R --vanilla' from the command line, the standard
matrix
> product is by default based on BLAS and multi-threaded on all cores
> available on the machine, c.f. following examples:
>
> n=10000
> p=1000
> q=5000
> A = matrix(runif(n*p),nrow=n, ncol=p)
> B = matrix(runif(p*q),nrow=p, ncol=q)
> C = A %*% B # multi-threaded matrix product
>
>
> However, the default behavior to use all available cores can be an
> issue, especially on shared computing resources or when the matrix
> product is used in parallelized section of codes (for instance with
> 'mclapply' from the 'parallel' package). For instance, the
default
> matrix product is single-threaded in R-3.3.2 (I ran a test on my
> machine), this new features will deeply affect the behavior of existing
> R packages that use other multi-threading solutions.
>
> Thanks to this stackoverflow question
>
(https://stackoverflow.com/questions/45794290/in-r-how-to-control-multi-threading-in-blas-parallel-matrix-product),
> I now know that it is possible to control the number of BLAS threads
> thanks to the package 'RhpcBLASctl'. However, being able to control
the
> number of threads should maybe not require to use an additional package.
>
> In addition, the doc 'matmult' does not mention this point, it
points to
> the 'options' doc page and especially the 'matprod'
section, in which
> the multi-threading is not discussed.
>
>
> Here is the results of the 'sessionInfo()' function on my machine
for
> R-3.4.1:
> R version 3.4.1 (2017-06-30)
> Platform: x86_64-pc-linux-gnu (64-bit)
> Running under: Ubuntu 16.04.3 LTS
>
> Matrix products: default
> BLAS: /usr/lib/openblas-base/libblas.so.3
> LAPACK: /usr/lib/libopenblasp-r0.2.18.so
>
> locale:
>    [1] LC_CTYPE=fr_FR.utf8       LC_NUMERIC=C
>    [3] LC_TIME=fr_FR.utf8        LC_COLLATE=fr_FR.utf8
>    [5] LC_MONETARY=fr_FR.utf8    LC_MESSAGES=fr_FR.utf8
>    [7] LC_PAPER=fr_FR.utf8       LC_NAME=C
>    [9] LC_ADDRESS=C              LC_TELEPHONE=C
> [11] LC_MEASUREMENT=fr_FR.utf8 LC_IDENTIFICATION=C
>
> attached base packages:
> [1] stats     graphics  grDevices utils     datasets  methods base
>
> loaded via a namespace (and not attached):
> [1] compiler_3.4.1
>
>
>
> and for R-3.3.2:
> R version 3.3.2 (2016-10-31)
> Platform: x86_64-pc-linux-gnu (64-bit)
> Running under: Ubuntu 16.04.3 LTS
>
> locale:
>    [1] LC_CTYPE=fr_FR.utf8       LC_NUMERIC=C
>    [3] LC_TIME=fr_FR.utf8        LC_COLLATE=fr_FR.utf8
>    [5] LC_MONETARY=fr_FR.utf8    LC_MESSAGES=fr_FR.utf8
>    [7] LC_PAPER=fr_FR.utf8       LC_NAME=C
>    [9] LC_ADDRESS=C              LC_TELEPHONE=C
> [11] LC_MEASUREMENT=fr_FR.utf8 LC_IDENTIFICATION=C
>
> attached base packages:
> [1] stats     graphics  grDevices utils     datasets  methods base
>
>
> Thanks in advance,
> Best regards
> ||
>

Many, though not all, threaded BLAS honor the OpenMP environment
variables; you might see whether setting OMP_THEAD_LIMIT or
OMP_NUM_THREADS does what you want.

Best,

luke

On Mon, 21 Aug 2017, Tomas Kalibera wrote:
> Hi Ghislain,
>
> I think you might be comparing two versions of R with different BLAS 
> implementations, one that is single threaded (is your 3.3.2 used with 
> reference blas?) and one that is multi threaded (3.4.1 with openblas).
Could
> you check with "perf"? E.g. run your benchmark with "perf
record" in both
> cases and you should see the names of the hot BLAS functions and this
should
> reveal the BLAS implementation (look for dgemm).
>
> In Ubuntu, if you install R from the package system, whenever you run it it
> will use the BLAS currently installed via the package system. However if
you
> build R from source on Ubuntu, by default, it will use the reference BLAS 
> which is distributed with R. Section "Linear algebra" of "R
Installation and
> Administration" has details on how to build R with different
BLAS/LAPACK
> implementations.
>
> Sadly there is no standard way to specify the number of BLAS worker
threads.
> RhpcBLASctl has specific code for several existing implementations, but R 
> itself does not attempt to control BLAS multi threading in any way. It is 
> expected the user/system administrator will configure their BLAS 
> implementation of choice to use the number of threads they need. A similar 
> problem exists in other internally multi-threaded third-party libraries,
used
> by packages - R cannot control how many threads they run.
>
> Best
> Tomas
>
> On 08/21/2017 02:55 PM, Ghislain Durif wrote:
>> Dear R Core Team,
>> 
>> I wish to report what can be viewed as a bug or at least a strange
>> behavior in R-3.4.1. I ask my question here (as recommended on
>> https://www.r-project.org/bugs.html) since I am not member of the
R's
>> Bugzilla.
>> 
>> When running 'R --vanilla' from the command line, the standard
matrix
>> product is by default based on BLAS and multi-threaded on all cores
>> available on the machine, c.f. following examples:
>> 
>> n=10000
>> p=1000
>> q=5000
>> A = matrix(runif(n*p),nrow=n, ncol=p)
>> B = matrix(runif(p*q),nrow=p, ncol=q)
>> C = A %*% B # multi-threaded matrix product
>> 
>> 
>> However, the default behavior to use all available cores can be an
>> issue, especially on shared computing resources or when the matrix
>> product is used in parallelized section of codes (for instance with
>> 'mclapply' from the 'parallel' package). For instance,
the default
>> matrix product is single-threaded in R-3.3.2 (I ran a test on my
>> machine), this new features will deeply affect the behavior of existing
>> R packages that use other multi-threading solutions.
>> 
>> Thanks to this stackoverflow question
>>
(https://stackoverflow.com/questions/45794290/in-r-how-to-control-multi-threading-in-blas-parallel-matrix-product),
>> I now know that it is possible to control the number of BLAS threads
>> thanks to the package 'RhpcBLASctl'. However, being able to
control the
>> number of threads should maybe not require to use an additional
package.
>> 
>> In addition, the doc 'matmult' does not mention this point, it
points to
>> the 'options' doc page and especially the 'matprod'
section, in which
>> the multi-threading is not discussed.
>> 
>> 
>> Here is the results of the 'sessionInfo()' function on my
machine for
>> R-3.4.1:
>> R version 3.4.1 (2017-06-30)
>> Platform: x86_64-pc-linux-gnu (64-bit)
>> Running under: Ubuntu 16.04.3 LTS
>> 
>> Matrix products: default
>> BLAS: /usr/lib/openblas-base/libblas.so.3
>> LAPACK: /usr/lib/libopenblasp-r0.2.18.so
>> 
>> locale:
>>    [1] LC_CTYPE=fr_FR.utf8       LC_NUMERIC=C
>>    [3] LC_TIME=fr_FR.utf8        LC_COLLATE=fr_FR.utf8
>>    [5] LC_MONETARY=fr_FR.utf8    LC_MESSAGES=fr_FR.utf8
>>    [7] LC_PAPER=fr_FR.utf8       LC_NAME=C
>>    [9] LC_ADDRESS=C              LC_TELEPHONE=C
>> [11] LC_MEASUREMENT=fr_FR.utf8 LC_IDENTIFICATION=C
>> 
>> attached base packages:
>> [1] stats     graphics  grDevices utils     datasets  methods base
>> 
>> loaded via a namespace (and not attached):
>> [1] compiler_3.4.1
>> 
>> 
>> 
>> and for R-3.3.2:
>> R version 3.3.2 (2016-10-31)
>> Platform: x86_64-pc-linux-gnu (64-bit)
>> Running under: Ubuntu 16.04.3 LTS
>> 
>> locale:
>>    [1] LC_CTYPE=fr_FR.utf8       LC_NUMERIC=C
>>    [3] LC_TIME=fr_FR.utf8        LC_COLLATE=fr_FR.utf8
>>    [5] LC_MONETARY=fr_FR.utf8    LC_MESSAGES=fr_FR.utf8
>>    [7] LC_PAPER=fr_FR.utf8       LC_NAME=C
>>    [9] LC_ADDRESS=C              LC_TELEPHONE=C
>> [11] LC_MEASUREMENT=fr_FR.utf8 LC_IDENTIFICATION=C
>> 
>> attached base packages:
>> [1] stats     graphics  grDevices utils     datasets  methods base
>> 
>> 
>> Thanks in advance,
>> Best regards
>> ||
>> 
>
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
>
-- 
Luke Tierney
Ralph E. Wareham Professor of Mathematical Sciences
University of Iowa                  Phone:             319-335-3386
Department of Statistics and        Fax:               319-335-3017
    Actuarial Science
241 Schaeffer Hall                  email:   luke-tierney at uiowa.edu
Iowa City, IA 52242                 WWW:  http://www.stat.uiowa.edu

Hi Tomas,

Thanks for your answer.

Indeed, I checked and my R-3.4.1 installed from the ubuntu repository 
use 'libopenblasp-r0.2.18.so' while my R-3.3.2 that I did compiled on my
machine use 'libRblas.so' which explain the difference of behavior.

I will use RhpcBLASctl to avoid issue when combining matrix product and 
other multi-threading package.

Maybe this point regarding multi-threading with BLAS could be added in 
the R doc.

Thanks again,
Best,

Ghislain

Ghislain Durif
--------------------------
Research engineer THOTH TEAM
INRIA Grenoble Alpes (France)

Le 21/08/2017 ? 15:53, Tomas Kalibera a ?crit :
> Hi Ghislain,
>
> I think you might be comparing two versions of R with different BLAS 
> implementations, one that is single threaded (is your 3.3.2 used with 
> reference blas?) and one that is multi threaded (3.4.1 with openblas). 
> Could you check with "perf"? E.g. run your benchmark with
"perf
> record" in both cases and you should see the names of the hot BLAS 
> functions and this should reveal the BLAS implementation (look for 
> dgemm).
>
> In Ubuntu, if you install R from the package system, whenever you run 
> it it will use the BLAS currently installed via the package system. 
> However if you build R from source on Ubuntu, by default, it will use 
> the reference BLAS which is distributed with R. Section "Linear 
> algebra" of "R Installation and Administration" has details
on how to
> build R with different BLAS/LAPACK implementations.
>
> Sadly there is no standard way to specify the number of BLAS worker 
> threads. RhpcBLASctl has specific code for several existing 
> implementations, but R itself does not attempt to control BLAS multi 
> threading in any way. It is expected the user/system administrator 
> will configure their BLAS implementation of choice to use the number 
> of threads they need. A similar problem exists in other internally 
> multi-threaded third-party libraries, used by packages - R cannot 
> control how many threads they run.
>
> Best
> Tomas
>
> On 08/21/2017 02:55 PM, Ghislain Durif wrote:
>> Dear R Core Team,
>>
>> I wish to report what can be viewed as a bug or at least a strange
>> behavior in R-3.4.1. I ask my question here (as recommended on
>> https://www.r-project.org/bugs.html) since I am not member of the
R's
>> Bugzilla.
>>
>> When running 'R --vanilla' from the command line, the standard
matrix
>> product is by default based on BLAS and multi-threaded on all cores
>> available on the machine, c.f. following examples:
>>
>> n=10000
>> p=1000
>> q=5000
>> A = matrix(runif(n*p),nrow=n, ncol=p)
>> B = matrix(runif(p*q),nrow=p, ncol=q)
>> C = A %*% B # multi-threaded matrix product
>>
>>
>> However, the default behavior to use all available cores can be an
>> issue, especially on shared computing resources or when the matrix
>> product is used in parallelized section of codes (for instance with
>> 'mclapply' from the 'parallel' package). For instance,
the default
>> matrix product is single-threaded in R-3.3.2 (I ran a test on my
>> machine), this new features will deeply affect the behavior of existing
>> R packages that use other multi-threading solutions.
>>
>> Thanks to this stackoverflow question
>>
(https://stackoverflow.com/questions/45794290/in-r-how-to-control-multi-threading-in-blas-parallel-matrix-product),
>>
>> I now know that it is possible to control the number of BLAS threads
>> thanks to the package 'RhpcBLASctl'. However, being able to
control the
>> number of threads should maybe not require to use an additional
package.
>>
>> In addition, the doc 'matmult' does not mention this point, it
points to
>> the 'options' doc page and especially the 'matprod'
section, in which
>> the multi-threading is not discussed.
>>
>>
>> Here is the results of the 'sessionInfo()' function on my
machine for
>> R-3.4.1:
>> R version 3.4.1 (2017-06-30)
>> Platform: x86_64-pc-linux-gnu (64-bit)
>> Running under: Ubuntu 16.04.3 LTS
>>
>> Matrix products: default
>> BLAS: /usr/lib/openblas-base/libblas.so.3
>> LAPACK: /usr/lib/libopenblasp-r0.2.18.so
>>
>> locale:
>>    [1] LC_CTYPE=fr_FR.utf8       LC_NUMERIC=C
>>    [3] LC_TIME=fr_FR.utf8        LC_COLLATE=fr_FR.utf8
>>    [5] LC_MONETARY=fr_FR.utf8    LC_MESSAGES=fr_FR.utf8
>>    [7] LC_PAPER=fr_FR.utf8       LC_NAME=C
>>    [9] LC_ADDRESS=C              LC_TELEPHONE=C
>> [11] LC_MEASUREMENT=fr_FR.utf8 LC_IDENTIFICATION=C
>>
>> attached base packages:
>> [1] stats     graphics  grDevices utils     datasets  methods base
>>
>> loaded via a namespace (and not attached):
>> [1] compiler_3.4.1
>>
>>
>>
>> and for R-3.3.2:
>> R version 3.3.2 (2016-10-31)
>> Platform: x86_64-pc-linux-gnu (64-bit)
>> Running under: Ubuntu 16.04.3 LTS
>>
>> locale:
>>    [1] LC_CTYPE=fr_FR.utf8       LC_NUMERIC=C
>>    [3] LC_TIME=fr_FR.utf8        LC_COLLATE=fr_FR.utf8
>>    [5] LC_MONETARY=fr_FR.utf8    LC_MESSAGES=fr_FR.utf8
>>    [7] LC_PAPER=fr_FR.utf8       LC_NAME=C
>>    [9] LC_ADDRESS=C              LC_TELEPHONE=C
>> [11] LC_MEASUREMENT=fr_FR.utf8 LC_IDENTIFICATION=C
>>
>> attached base packages:
>> [1] stats     graphics  grDevices utils     datasets  methods base
>>
>>
>> Thanks in advance,
>> Best regards
>> ||
>>
>