[Rcpp-devel] RcppArmadillo Memory Leak?

Dirk Eddelbuettel edd at debian.org
Wed Nov 20 19:51:11 CET 2013


On 20 November 2013 at 19:24, Alessandro Vitale wrote:
| I've done the new test (dense matrix with just Rcpp NumericMatrix,  no RcppArmadillo) and in this case there is no leak.
| 
| This is the code I used
| 
| // [[Rcpp::export]]
| NumericMatrix dense_numericMatrix(NumericMatrix A){
|     A(0,0) = 1;
|     A(1,0) = 2;
|     return A ;
| }
| 
| 
| testMatrix <- matrix(data=0,nrow=1e3*5,ncol=1e3) # 40MB Dense Matrix
| for(i in 1:5){
|   a <- dense_numericMatrix(testMatrix)	
| }
| 
| and this is valgrind output
| 
| ==28325== LEAK SUMMARY:
| ==28325==    definitely lost: 36 bytes in 2 blocks
| ==28325==    indirectly lost: 0 bytes in 0 blocks
| ==28325==      possibly lost: 0 bytes in 0 blocks
| ==28325==    still reachable: 74,382,445 bytes in 20,450 blocks
| ==28325==         suppressed: 21,484 bytes in 82 blocks
| ==28325== 
| ==28325== For counts of detected and suppressed errors, rerun with: -v
| ==28325== ERROR SUMMARY: 2 errors from 2 contexts (suppressed: 10 from 10)

Cool. Thanks for doing that.
 
| Can I support in some other ways in finding what's causing the problem in RcppArmadillo sparse matrix?

We'd have to dig through the transition from the arma::sp_mat to what is
returned to R.  

You could possibly start with the posts I had on the Rcpp Gallery as they
preceded some of the other work.

Thanks for doing that -- really appreciate it.

Dirk

 
| Alessandro
| 
| 
| On 20 Nov, 2013, at 18:36 , Dirk Eddelbuettel wrote:
| 
| > 
| > On 18 November 2013 at 18:16, Alessandro Vitale wrote:
| > | Dirk, I've run the code a couple times with valgrind, one time calling sparse 5 times, the other calling sparse 10 times.
| > | Below is the main code and valgrind output.
| > | What do you think, could it be an Armadillo / RcppArmadillo leak?
| > 
| > Could be.  For comparison, what happens when you use a regular matrix (ie
| > Rcpp::NumericMatrix) and check with valgrind?
| > 
| > Dirk
| > 
| > 
| > | testMatrix <- Matrix(data=0,nrow=1e7,ncol=1e7,sparse=TRUE)
| > | 
| > | for(i in 1:5){
| > | 	a <- sparse(testMatrix)	
| > | }
| > | 
| > | ==22103== 160,000,032 bytes in 2 blocks are possibly lost in loss record 4,034 of 4,035
| > | ==22103==    at 0xC658: malloc (vg_replace_malloc.c:295)
| > | ==22103==    by 0x8BD503A: arma::SpMat<double>::init(unsigned long long, unsigned long long) (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/Rtmp8Krm0b/sourcecpp_565778edd300/sourceCpp_27814.so)
| > | ==22103==    by 0x8BD6690: Rcpp::traits::Exporter<arma::SpMat<double> >::get() (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/Rtmp8Krm0b/sourcecpp_565778edd300/sourceCpp_27814.so)
| > | ==22103==    by 0x8BD6B94: arma::SpMat<double> Rcpp::internal::as<arma::SpMat<double> >(SEXPREC*, Rcpp::traits::r_type_generic_tag) (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/Rtmp8Krm0b/sourcecpp_565778edd300/sourceCpp_27814.so)
| > | ==22103==    by 0x8BD4B37: sourceCpp_12484_sparse (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/Rtmp8Krm0b/sourcecpp_565778edd300/sourceCpp_27814.so)
| > | ==22103==    by 0x86D84: do_dotcall (dotcode.c:579)
| > | ==22103==    by 0xAEEF3: Rf_eval (eval.c:635)
| > | ==22103==    by 0xB8E6D: Rf_applyClosure (eval.c:1003)
| > | ==22103==    by 0xAEF93: Rf_eval (eval.c:654)
| > | ==22103==    by 0xBB1DE: do_set (eval.c:1871)
| > | ==22103==    by 0xAED66: Rf_eval (eval.c:607)
| > | ==22103==    by 0xBB53F: do_begin (eval.c:1557)
| > | ==22103== 
| > | ==22103== 240,000,048 bytes in 3 blocks are definitely lost in loss record 4,035 of 4,035
| > | ==22103==    at 0xC658: malloc (vg_replace_malloc.c:295)
| > | ==22103==    by 0x8BD503A: arma::SpMat<double>::init(unsigned long long, unsigned long long) (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/Rtmp8Krm0b/sourcecpp_565778edd300/sourceCpp_27814.so)
| > | ==22103==    by 0x8BD6690: Rcpp::traits::Exporter<arma::SpMat<double> >::get() (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/Rtmp8Krm0b/sourcecpp_565778edd300/sourceCpp_27814.so)
| > | ==22103==    by 0x8BD6B94: arma::SpMat<double> Rcpp::internal::as<arma::SpMat<double> >(SEXPREC*, Rcpp::traits::r_type_generic_tag) (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/Rtmp8Krm0b/sourcecpp_565778edd300/sourceCpp_27814.so)
| > | ==22103==    by 0x8BD4B37: sourceCpp_12484_sparse (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/Rtmp8Krm0b/sourcecpp_565778edd300/sourceCpp_27814.so)
| > | ==22103==    by 0x86D84: do_dotcall (dotcode.c:579)
| > | ==22103==    by 0xAEEF3: Rf_eval (eval.c:635)
| > | ==22103==    by 0xB8E6D: Rf_applyClosure (eval.c:1003)
| > | ==22103==    by 0xAEF93: Rf_eval (eval.c:654)
| > | ==22103==    by 0xBB1DE: do_set (eval.c:1871)
| > | ==22103==    by 0xAED66: Rf_eval (eval.c:607)
| > | ==22103==    by 0xBB53F: do_begin (eval.c:1557)
| > | ==22103== 
| > | ==22103== LEAK SUMMARY:
| > | ==22103==    definitely lost: 240,018,516 bytes in 14 blocks
| > | ==22103==    indirectly lost: 0 bytes in 0 blocks
| > | ==22103==      possibly lost: 160,010,424 bytes in 4 blocks
| > | ==22103==    still reachable: 210,051,491 bytes in 52,853 blocks
| > | ==22103==         suppressed: 25,144 bytes in 94 blocks
| > | ==22103== 
| > | ==22103== For counts of detected and suppressed errors, rerun with: -v
| > | ==22103== ERROR SUMMARY: 8 errors from 8 contexts (suppressed: 10 from 10)
| > | 
| > | 
| > | 
| > | testMatrix <- Matrix(data=0,nrow=1e7,ncol=1e7,sparse=TRUE)
| > | 
| > | for(i in 1:10){
| > | 	a <- sparse(testMatrix)	
| > | }
| > | 
| > | ==22148== 
| > | ==22148== 320,000,064 bytes in 4 blocks are possibly lost in loss record 4,039 of 4,040
| > | ==22148==    at 0xC658: malloc (vg_replace_malloc.c:295)
| > | ==22148==    by 0x8BD503A: arma::SpMat<double>::init(unsigned long long, unsigned long long) (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/RtmpLZuqjl/sourcecpp_56846fe5a651/sourceCpp_83305.so)
| > | ==22148==    by 0x8BD6690: Rcpp::traits::Exporter<arma::SpMat<double> >::get() (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/RtmpLZuqjl/sourcecpp_56846fe5a651/sourceCpp_83305.so)
| > | ==22148==    by 0x8BD6B94: arma::SpMat<double> Rcpp::internal::as<arma::SpMat<double> >(SEXPREC*, Rcpp::traits::r_type_generic_tag) (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/RtmpLZuqjl/sourcecpp_56846fe5a651/sourceCpp_83305.so)
| > | ==22148==    by 0x8BD4B37: sourceCpp_93626_sparse (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/RtmpLZuqjl/sourcecpp_56846fe5a651/sourceCpp_83305.so)
| > | ==22148==    by 0x86D84: do_dotcall (dotcode.c:579)
| > | ==22148==    by 0xAEEF3: Rf_eval (eval.c:635)
| > | ==22148==    by 0xB8E6D: Rf_applyClosure (eval.c:1003)
| > | ==22148==    by 0xAEF93: Rf_eval (eval.c:654)
| > | ==22148==    by 0xBB1DE: do_set (eval.c:1871)
| > | ==22148==    by 0xAED66: Rf_eval (eval.c:607)
| > | ==22148==    by 0xBB53F: do_begin (eval.c:1557)
| > | ==22148== 
| > | ==22148== 480,000,096 bytes in 6 blocks are definitely lost in loss record 4,040 of 4,040
| > | ==22148==    at 0xC658: malloc (vg_replace_malloc.c:295)
| > | ==22148==    by 0x8BD503A: arma::SpMat<double>::init(unsigned long long, unsigned long long) (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/RtmpLZuqjl/sourcecpp_56846fe5a651/sourceCpp_83305.so)
| > | ==22148==    by 0x8BD6690: Rcpp::traits::Exporter<arma::SpMat<double> >::get() (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/RtmpLZuqjl/sourcecpp_56846fe5a651/sourceCpp_83305.so)
| > | ==22148==    by 0x8BD6B94: arma::SpMat<double> Rcpp::internal::as<arma::SpMat<double> >(SEXPREC*, Rcpp::traits::r_type_generic_tag) (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/RtmpLZuqjl/sourcecpp_56846fe5a651/sourceCpp_83305.so)
| > | ==22148==    by 0x8BD4B37: sourceCpp_93626_sparse (in /private/var/folders/2h/s9877ww97794_0fg6vrtm54c0000gn/T/RtmpLZuqjl/sourcecpp_56846fe5a651/sourceCpp_83305.so)
| > | ==22148==    by 0x86D84: do_dotcall (dotcode.c:579)
| > | ==22148==    by 0xAEEF3: Rf_eval (eval.c:635)
| > | ==22148==    by 0xB8E6D: Rf_applyClosure (eval.c:1003)
| > | ==22148==    by 0xAEF93: Rf_eval (eval.c:654)
| > | ==22148==    by 0xBB1DE: do_set (eval.c:1871)
| > | ==22148==    by 0xAED66: Rf_eval (eval.c:607)
| > | ==22148==    by 0xBB53F: do_begin (eval.c:1557)
| > | ==22148== 
| > | ==22148== LEAK SUMMARY:
| > | ==22148==    definitely lost: 480,041,092 bytes in 28 blocks
| > | ==22148==    indirectly lost: 0 bytes in 0 blocks
| > | ==22148==      possibly lost: 320,008,408 bytes in 5 blocks
| > | ==22148==    still reachable: 217,092,931 bytes in 52,888 blocks
| > | ==22148==         suppressed: 25,144 bytes in 94 blocks
| > | ==22148== 
| > | ==22148== For counts of detected and suppressed errors, rerun with: -v
| > | ==22148== ERROR SUMMARY: 7 errors from 7 contexts (suppressed: 10 from 10)
| > | 
| > | 
| > | 
| > | 
| > | 
| > | On 18 Nov, 2013, at 15:59 , Alessandro Vitale wrote:
| > | 
| > | > Thanks Dirk and Hadley.
| > | > 
| > | > Dirk, asap I will try your b) and c) suggestions. While for suggestion a), I didn't really understand how to implement it.
| > | > 
| > | > Hadley, I am determining rsession process memory looking at osx activity monitor utility, under Real Mem column.
| > | > This is what I get (I am using your own mem() function to determine r memory usage)
| > | > 
| > | > testMatrix <- Matrix(data=0,nrow=1e7,ncol=1e7,sparse=TRUE)
| > | > gc()
| > | > mem() # [1] 123.3, rsession Real Mem 203.1 MB
| > | > sparse(testMatrix)
| > | > mem() # [1] 161.5
| > | > gc()
| > | > mem() # [1] 123.4, rsession Real Mem 507.8 MB
| > | > sparse(testMatrix)
| > | > mem() # [1] 161.5
| > | > gc()
| > | > mem() # [1] 123.4,  rsession Real Mem 585.0 MB
| > | > sparse(testMatrix)
| > | > mem() # [1] 161.5
| > | > gc()
| > | > mem() # [1] 123.4,  rsession Real Mem 661.3 MB
| > | > 
| > | > and so on.
| > | > 
| > | > If I call the function too many times (in fact I discovered it while using benchmark() ), it fills up the memory and the computer gets totally unresponsive.
| > | > 
| > | > 
| > | > 
| > | > 
| > | > 
| > | > On 18 Nov, 2013, at 15:24 , Hadley Wickham wrote:
| > | > 
| > | >>> After every call, the memory from gc() is stable, while rsession process
| > | >>> memory keeps growing of around 100Mb at each call and doesn't decrease after
| > | >>> the garbage collection.
| > | >> 
| > | >> You don't mention how you're determining the size of the rsession
| > | >> process, but you shouldn't necessarily expect it to decrease after
| > | >> calling gc() - there may be memory fragmentation, but also the OS is
| > | >> lazy about reclaiming memory from processes, so it may only be
| > | >> returned to the OS when it actually needs it.
| > | >> 
| > | >> Hadley
| > | >> 
| > | >> -- 
| > | >> Chief Scientist, RStudio
| > | >> http://had.co.nz/
| > | > 
| > | 
| > | _______________________________________________
| > | Rcpp-devel mailing list
| > | Rcpp-devel at lists.r-forge.r-project.org
| > | https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/rcpp-devel
| > 
| > -- 
| > Dirk Eddelbuettel | edd at debian.org | http://dirk.eddelbuettel.com
| 

-- 
Dirk Eddelbuettel | edd at debian.org | http://dirk.eddelbuettel.com


More information about the Rcpp-devel mailing list