[Rd] R 3.0.0 memory use
I did some benchmarking of data frame code, and it appears that R 3.0.0 is far worse than earlier versions of R in terms of how many large objects it allocates space for, for data frame operations - creation, subscripting, subscript replacement. For a data frame with n rows, it makes either 2 or 4 extra copies of all of: 8n bytes (e.g. double precision) 24n bytes 32n bytes E.g., for as.data.frame(numeric vector), instead of allocations totalling ~8n bytes, it allocates 33 times that much. Here, compare columns 3 and 5 (columns 2 and 4 are with the dataframe package). # Summary # R-2.14.2R-2.15.3R-3.0.0 # w/o withw/o withw/o # as.data.frame(y)3 1 1 1 5;4;4 # data.frame(y) 7 3 4 2 6;2;2 # data.frame(y, z)7 each 3 each 4 2 8;4;4 # as.data.frame(l)8 3 5 2 9;4;4 # data.frame(l) 13 5 8 3 12;4;4 # d$z - z3,2 1,1 3,1 2,1 7;4;4,1 # d[[z]] - z 4,3 1,1 3,1 2,1 7;4;4,1 # d[, z] - z 6,4,2 2,2,1 4,2,2 3,2,1 8;4;4,2,2 # d[z] - z 6,5,2 2,2,1 4,2,2 3,2,1 8;4;4,2,2 # d[z] - list(z=z) 6,3,2 2,2,1 4,2,2 3,2,1 8;4;4,2,2 # d[z] - Z #list(z=z) 6,2,2 2,1,1 4,1,2 3,1,1 8;4;4,1,2 # a - d[y] 2 1 2 1 6;4;4 # a - d[, y, drop=F] 2 1 2 1 6;4;4 # Where two numbers are given, they refer to: # (copies of the old data frame), # (copies of the new column) # A third number refers to numbers of # (copies made of an integer vector of row names) # For R 3.0.0, I'm getting astounding results - many more copies, # and also some copies of larger objects; in addition to the data # vectors of size 80K and 160K, also 240K and 320K. # Where three numbers are given in form a;c;d, they refer to # (copies of 80K; 240K; 320K) The benchmarks are at http://www.timhesterberg.net/r-packages/memory.R I'm using versions of R I installed from source on a Linux box, using e.g. ./configure --prefix=(my path) --enable-memory-profiling --with-readline=no make make install __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Re: [Rd] R 3.0.0 memory use
There were a couple of bug fixes to somewhat obscure compound assignment related bugs that required bumping up internal reference counts. It's possible that one or more of these are responsible. If so it is unavoidable for now, but it's worth finding out for sure. With some stripped down test examples it should be possible to identify when things changed. I won't have time to look for some time, but if someone else wanted to nail this down that would be useful. Best, luke On Sun, 14 Apr 2013, Tim Hesterberg wrote: I did some benchmarking of data frame code, and it appears that R 3.0.0 is far worse than earlier versions of R in terms of how many large objects it allocates space for, for data frame operations - creation, subscripting, subscript replacement. For a data frame with n rows, it makes either 2 or 4 extra copies of all of: 8n bytes (e.g. double precision) 24n bytes 32n bytes E.g., for as.data.frame(numeric vector), instead of allocations totalling ~8n bytes, it allocates 33 times that much. Here, compare columns 3 and 5 (columns 2 and 4 are with the dataframe package). # Summary # R-2.14.2R-2.15.3R-3.0.0 # w/o withw/o withw/o # as.data.frame(y)3 1 1 1 5;4;4 # data.frame(y) 7 3 4 2 6;2;2 # data.frame(y, z)7 each 3 each 4 2 8;4;4 # as.data.frame(l)8 3 5 2 9;4;4 # data.frame(l) 13 5 8 3 12;4;4 # d$z - z3,2 1,1 3,1 2,1 7;4;4,1 # d[[z]] - z 4,3 1,1 3,1 2,1 7;4;4,1 # d[, z] - z 6,4,2 2,2,1 4,2,2 3,2,1 8;4;4,2,2 # d[z] - z 6,5,2 2,2,1 4,2,2 3,2,1 8;4;4,2,2 # d[z] - list(z=z) 6,3,2 2,2,1 4,2,2 3,2,1 8;4;4,2,2 # d[z] - Z #list(z=z) 6,2,2 2,1,1 4,1,2 3,1,1 8;4;4,1,2 # a - d[y] 2 1 2 1 6;4;4 # a - d[, y, drop=F] 2 1 2 1 6;4;4 # Where two numbers are given, they refer to: # (copies of the old data frame), # (copies of the new column) # A third number refers to numbers of # (copies made of an integer vector of row names) # For R 3.0.0, I'm getting astounding results - many more copies, # and also some copies of larger objects; in addition to the data # vectors of size 80K and 160K, also 240K and 320K. # Where three numbers are given in form a;c;d, they refer to # (copies of 80K; 240K; 320K) The benchmarks are at http://www.timhesterberg.net/r-packages/memory.R I'm using versions of R I installed from source on a Linux box, using e.g. ./configure --prefix=(my path) --enable-memory-profiling --with-readline=no make make install __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel -- Luke Tierney Chair, Statistics and Actuarial Science Ralph E. Wareham Professor of Mathematical Sciences University of Iowa Phone: 319-335-3386 Department of Statistics andFax: 319-335-3017 Actuarial Science 241 Schaeffer Hall email: luke-tier...@uiowa.edu Iowa City, IA 52242 WWW: http://www.stat.uiowa.edu __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Re: [Rd] R 3.0.0 memory use
On 04/14/2013 07:11 PM, luke-tier...@uiowa.edu wrote: There were a couple of bug fixes to somewhat obscure compound assignment related bugs that required bumping up internal reference counts. It's possible that one or more of these are responsible. If so it is unavoidable for now, but it's worth finding out for sure. With some stripped down test examples it should be possible to identify when things changed. I won't have time to look for some time, but if someone else wanted to nail this down that would be useful. I can't quite tell from Tim's script what he's documenting. In R-2.15.3 I have Rprofmem(); Rprofmem(NULL); readLines(Rprofmem.out, warn=FALSE) character(0) (or sometimes [1] new page:new page:\Rprofmem\ ) whereas in R-3.0.0 Rprofmem(); Rprofmem(NULL); readLines(Rprofmem.out, warn=FALSE) [1] 320040 :80040 :240048 :320040 :80040 :240048 : I think these are the allocations Tim is seeing. They're from the parser (see below) rather than as.data.frame. For Tim's example y - 1:10^4 + 0.0 Rprofmem(); d - as.data.frame(y); Rprofmem(NULL); readLines(Rprofmem.out) [1] 320040 :80040 :240048 :320040 :80040 :240048 :80040 :\as.data.frame.numeric\ \as.data.frame\ [2] 320040 :80040 :240048 :320040 :80040 :240048 : only the allocation 80040 is from as.data.frame (from the call stack output). Under R -d gdb (gdb) b R_OutputStackTrace (gdb) r Rprofmem(); Rprofmem(NULL) Breakpoint 1, R_OutputStackTrace (file=0xbd43f0) at /home/mtmorgan/src/R-3-0-branch/src/main/memory.c:3434 3434 { (gdb) bt #0 R_OutputStackTrace (file=0xbd43f0) at /home/mtmorgan/src/R-3-0-branch/src/main/memory.c:3434 #1 0x7792ff83 in R_ReportAllocation (size=320040) at /home/mtmorgan/src/R-3-0-branch/src/main/memory.c:3456 #2 Rf_allocVector (type=13, length=8) at /home/mtmorgan/src/R-3-0-branch/src/main/memory.c:2478 #3 0x7790bedf in growData () at gram.y:3391 and the memory allocations are from these lines in the parser gram.y PROTECT( bigger = allocVector( INTSXP, data_size * DATA_ROWS ) ) ; PROTECT( biggertext = allocVector( STRSXP, data_size ) ); I'm not sure why these show up under R 3.0.0, though. $ R-2-15-branch/bin/R --version R version 2.15.3 Patched (2013-03-13 r62579) -- Security Blanket Copyright (C) 2013 The R Foundation for Statistical Computing ISBN 3-900051-07-0 Platform: x86_64-unknown-linux-gnu (64-bit) R-3-0-branch$ bin/R --version R version 3.0.0 Patched (2013-04-14 r62579) -- Masked Marvel Copyright (C) 2013 The R Foundation for Statistical Computing Platform: x86_64-unknown-linux-gnu (64-bit) Martin Best, luke On Sun, 14 Apr 2013, Tim Hesterberg wrote: I did some benchmarking of data frame code, and it appears that R 3.0.0 is far worse than earlier versions of R in terms of how many large objects it allocates space for, for data frame operations - creation, subscripting, subscript replacement. For a data frame with n rows, it makes either 2 or 4 extra copies of all of: 8n bytes (e.g. double precision) 24n bytes 32n bytes E.g., for as.data.frame(numeric vector), instead of allocations totalling ~8n bytes, it allocates 33 times that much. Here, compare columns 3 and 5 (columns 2 and 4 are with the dataframe package). # Summary # R-2.14.2R-2.15.3R-3.0.0 # w/o withw/o withw/o # as.data.frame(y)3 1 1 1 5;4;4 # data.frame(y) 7 3 4 2 6;2;2 # data.frame(y, z)7 each 3 each 4 2 8;4;4 # as.data.frame(l)8 3 5 2 9;4;4 # data.frame(l) 13 5 8 3 12;4;4 # d$z - z3,2 1,1 3,1 2,1 7;4;4,1 # d[[z]] - z 4,3 1,1 3,1 2,1 7;4;4,1 # d[, z] - z 6,4,2 2,2,1 4,2,2 3,2,1 8;4;4,2,2 # d[z] - z 6,5,2 2,2,1 4,2,2 3,2,1 8;4;4,2,2 # d[z] - list(z=z) 6,3,2 2,2,1 4,2,2 3,2,1 8;4;4,2,2 # d[z] - Z #list(z=z) 6,2,2 2,1,1 4,1,2 3,1,1 8;4;4,1,2 # a - d[y] 2 1 2 1 6;4;4 # a - d[, y, drop=F] 2 1 2 1 6;4;4 # Where two numbers are given, they refer to: # (copies of the old data frame), # (copies of the new column) # A third number refers to numbers of # (copies made of an integer vector of row names) # For R 3.0.0, I'm getting astounding results - many more copies, # and also some copies of larger objects; in addition to the data # vectors of size 80K and 160K, also 240K and 320K. # Where three numbers are given in form a;c;d, they refer to # (copies of 80K; 240K; 320K) The benchmarks are at http://www.timhesterberg.net/r-packages/memory.R I'm using versions of R I installed from source on a Linux box, using e.g. ./configure --prefix=(my
Re: [Rd] R 3.0.0 memory use
When I change the data set size, the extra allocations do not change in size. This supports Luke and Martin's diagnosis. The extra allocations are either 2 or 4 allocations each of size 80040 240048 320040 Details (you may skip): (Fresh session of R 3.0.0) y - 1:10^4 + 0.0 Rprofmem(temp.out, threshold = 10^4) d - as.data.frame(y) Rprofmem(NULL); system(cat temp.out) 320040 :80040 :240048 :320040 :80040 :240048 :80040 :as.data.frame.numeric as.data.frame 320040 :80040 :240048 :320040 :80040 :240048 : # Try increasing size by a factor of 10 y - 1:10^5 + 0.0 Rprofmem(temp.out, threshold = 10^4) d - as.data.frame(y) Rprofmem(NULL); system(cat temp.out) 320040 :80040 :240048 :320040 :80040 :240048 :800040 :as.data.frame.numeric as.data.frame 320040 :80040 :240048 :320040 :80040 :240048 : The number of allocations shown, of different sizes: 3.0.0 3.0.0 2.15.3 2.15.3 first second first second 240048 4 4 0 0 320040 4 4 0 0 80040 5 4 1 0 800040 0 1 0 1 So it looks like both R 2.15.3 and R 3.0.0 are making one copy of the data, plus extra allocations. (Fresh session of R 2.15.3) y - 1:10^4 + 0.0 Rprofmem(temp.out, threshold = 10^4) d - as.data.frame(y) Rprofmem(NULL); system(cat temp.out) 80040 :as.data.frame.numeric as.data.frame # Increase size by factor of 10 y - 1:10^5 + 0.0 Rprofmem(temp.out, threshold = 10^4) d - as.data.frame(y) Rprofmem(NULL); system(cat temp.out) 800040 :as.data.frame.numeric as.data.frame On Sun, 14 Apr 2013 19:15:45 -0700 Martin Morgan mtmor...@fhcrc.org wrote: On 04/14/2013 07:11 PM, luke-tier...@uiowa.edu wrote: There were a couple of bug fixes to somewhat obscure compound assignment related bugs that required bumping up internal reference counts. It's possible that one or more of these are responsible. If so it is unavoidable for now, but it's worth finding out for sure. With some stripped down test examples it should be possible to identify when things changed. I won't have time to look for some time, but if someone else wanted to nail this down that would be useful. I can't quite tell from Tim's script what he's documenting. In R-2.15.3 I have Rprofmem(); Rprofmem(NULL); readLines(Rprofmem.out, warn=FALSE) character(0) (or sometimes [1] new page:new page:\Rprofmem\ ) whereas in R-3.0.0 Rprofmem(); Rprofmem(NULL); readLines(Rprofmem.out, warn=FALSE) [1] 320040 :80040 :240048 :320040 :80040 :240048 : I think these are the allocations Tim is seeing. They're from the parser (see below) rather than as.data.frame. For Tim's example y - 1:10^4 + 0.0 Rprofmem(); d - as.data.frame(y); Rprofmem(NULL); readLines(Rprofmem.out) [1] 320040 :80040 :240048 :320040 :80040 :240048 :80040 :\as.data.frame.numeric\ \as.data.frame\ [2] 320040 :80040 :240048 :320040 :80040 :240048 : only the allocation 80040 is from as.data.frame (from the call stack output). Under R -d gdb (gdb) b R_OutputStackTrace (gdb) r Rprofmem(); Rprofmem(NULL) Breakpoint 1, R_OutputStackTrace (file=0xbd43f0) at /home/mtmorgan/src/R-3-0-branch/src/main/memory.c:3434 3434{ (gdb) bt #0 R_OutputStackTrace (file=0xbd43f0) at /home/mtmorgan/src/R-3-0-branch/src/main/memory.c:3434 #1 0x7792ff83 in R_ReportAllocation (size=320040) at /home/mtmorgan/src/R-3-0-branch/src/main/memory.c:3456 #2 Rf_allocVector (type=13, length=8) at /home/mtmorgan/src/R-3-0-branch/src/main/memory.c:2478 #3 0x7790bedf in growData () at gram.y:3391 and the memory allocations are from these lines in the parser gram.y PROTECT( bigger = allocVector( INTSXP, data_size * DATA_ROWS ) ) ; PROTECT( biggertext = allocVector( STRSXP, data_size ) ); I'm not sure why these show up under R 3.0.0, though. $ R-2-15-branch/bin/R --version R version 2.15.3 Patched (2013-03-13 r62579) -- Security Blanket Copyright (C) 2013 The R Foundation for Statistical Computing ISBN 3-900051-07-0 Platform: x86_64-unknown-linux-gnu (64-bit) R-3-0-branch$ bin/R --version R version 3.0.0 Patched (2013-04-14 r62579) -- Masked Marvel Copyright (C) 2013 The R Foundation for Statistical Computing Platform: x86_64-unknown-linux-gnu (64-bit) Martin Best, luke On Sun, 14 Apr 2013, Tim Hesterberg wrote: I did some benchmarking of data frame code, and it appears that R 3.0.0 is far worse than earlier versions of R in terms of how many large objects it allocates space for, for data frame operations - creation, subscripting, subscript replacement. For a data frame with n rows, it makes either 2 or 4 extra copies of all of: 8n bytes (e.g. double precision) 24n bytes 32n bytes E.g., for as.data.frame(numeric vector), instead of allocations totalling ~8n bytes, it allocates 33 times that much. Here, compare columns 3 and 5 (columns 2 and 4 are with the dataframe package). # Summary #