[R] memory use of copies

Ross Boylan ross at biostat.ucsf.edu
Wed Jan 29 00:53:09 CET 2014


Thank you for a very thorough analysis.  It seems whether or not an
operation makes a full copy really depends on the specific operation,
and that it is not safe to assume that because I know something is
unchanged there will be no copy.  For example, in your last case only
one element of a list was modified, but all the list elements got new
memory.

BTW, one reason I got into this, aside from wanting to save memory, is
that I found my code was spending a lot of time in areas that probably
involved getting new memory.  So it mattered for speed too.

Ross

On Mon, 2014-01-27 at 06:33 -0800, Martin Morgan wrote:
> Hi Ross --
> 
> On 01/23/2014 05:53 PM, Ross Boylan wrote:
> > [Apologies if a duplicate; we are having mail problems.]
> >
> > I am trying to understand the circumstances under which R makes a copy
> > of an object, as opposed to simply referring to it.  I'm talking about
> > what goes on under the hood, not the user semantics.  I'm doing things
> > that take a lot of memory, and am trying to minimize my use.
> >
> > I thought that R was clever so that copies were created lazily.  For
> > example, if a is matrix, then
> > b <- a
> > b & a referred to to the same object underneath, so that a complete
> > duplicate (deep copy) wasn't made until it was necessary, e.g.,
> > b[3, 1] <- 4
> > would duplicate the contents of a to b, and then overwrite them.
> 
> Compiling your R with --enable-memory-profiling gives access to the tracemem() 
> function, showing that your understanding above is correct
> 
>  > b = matrix(0, 3, 2)
>  > tracemem(b)
> [1] "<0x7054020>"
>  > a = b        ## no copy
>  > b[3, 1] = 2  ## copy
> tracemem[0x7054020 -> 0x7053fc8]:
>  > b = matrix(0, 3, 2)
>  > tracemem(b)
>  > tracemem(b)
> [1] "<0x680e258>"
>  > b[3, 1] = 2  ## no copy
>  >
> 
> The same is apparent using .Internal(inspect()), where the first information 
> @7053ec0 is the address of the data. The other relevant part is the 'NAM()' 
> field, which indicates whether there are 0, 1 or (have been) at least 2 symbols 
> referring to the data. NAM() increments from 1 (no duplication on modify 
> required) on original creation to 2 when a = b (duplicate on modify)
> 
>  > b = matrix(0, 3, 2)
>  > .Internal(inspect(b))
> @7053ec0 14 REALSXP g0c4 [NAM(1),ATT] (len=6, tl=0) 0,0,0,0,0,...
> ATTRIB:
>    @7057528 02 LISTSXP g0c0 []
>      TAG: @21c5fb8 01 SYMSXP g0c0 [LCK,gp=0x4000] "dim" (has value)
>      @7056858 13 INTSXP g0c1 [NAM(2)] (len=2, tl=0) 3,2
>  > b[3, 1] = 2
>  > .Internal(inspect(b))
> @7053ec0 14 REALSXP g0c4 [NAM(1),ATT] (len=6, tl=0) 0,0,2,0,0,...
> ATTRIB:
>    @7057528 02 LISTSXP g0c0 []
>      TAG: @21c5fb8 01 SYMSXP g0c0 [LCK,gp=0x4000] "dim" (has value)
>      @7056858 13 INTSXP g0c1 [NAM(2)] (len=2, tl=0) 3,2
>  > a = b
>  > .Internal(inspect(b))      ## data address unchanced
> @7053ec0 14 REALSXP g0c4 [NAM(2),ATT] (len=6, tl=0) 0,0,0,0,0,...
> ATTRIB:
>    @7057528 02 LISTSXP g0c0 []
>      TAG: @21c5fb8 01 SYMSXP g0c0 [LCK,gp=0x4000] "dim" (has value)
>      @7056858 13 INTSXP g0c1 [NAM(2)] (len=2, tl=0) 3,2
>  > b[3, 1] = 2
>  > .Internal(inspect(b))      ## data address changed
> @7232910 14 REALSXP g0c4 [NAM(1),ATT] (len=6, tl=0) 0,0,2,0,0,...
> ATTRIB:
>    @7239d28 02 LISTSXP g0c0 []
>      TAG: @21c5fb8 01 SYMSXP g0c0 [LCK,gp=0x4000] "dim" (has value)
>      @7237b48 13 INTSXP g0c1 [NAM(2)] (len=2, tl=0) 3,2
> 
> 
> >
> > The following log, from R 3.0.1, does not seem to act that way; I get
> > the same amount of memory used whether I copy the same object repeatedly
> > or create new objects of the same size.
> >
> > Can anyone explain what is going on?  Am I just wrong that copies are
> > initially shallow?  Or perhaps that behavior only applies for function
> > arguments?  Or doesn't apply for class slots or reference class
> > variables?
> >
> >   > foo <- setRefClass("foo", fields=list(x="ANY"))
> >   > bar <- setClass("bar", slots=c("x"))
> 
> using the approach above, we can see that creating an S4 or reference object in 
> the way you've indicated (validity checks or other initialization might change 
> this) does not copy the data although it is marked for duplication
> 
>  > x = 1:2; .Internal(inspect(x))
> @7553868 13 INTSXP g0c1 [NAM(1)] (len=2, tl=0) 1,2
>  > .Internal(inspect(foo(x=x)$x))
> @7553868 13 INTSXP g0c1 [NAM(2)] (len=2, tl=0) 1,2
>  > .Internal(inspect(bar(x=x)@x))
> @7553868 13 INTSXP g0c1 [NAM(2)] (len=2, tl=0) 1,2
> 
> On the other hand, lapply is creating copies
> 
>  > x = 1:2; .Internal(inspect(x))
> @757b5a8 13 INTSXP g0c1 [NAM(1)] (len=2, tl=0) 1,2
>  > .Internal(inspect(lapply(1:2, function(i) x)))
> @7551f88 19 VECSXP g0c2 [] (len=2, tl=0)
>    @757b428 13 INTSXP g0c1 [] (len=2, tl=0) 1,2
>    @757b3f8 13 INTSXP g0c1 [] (len=2, tl=0) 1,2
> 
> One can construct a list without copies
> 
>  > x = 1:2; .Internal(inspect(x))
> @7677c18 13 INTSXP g0c1 [NAM(1)] (len=2, tl=0) 1,2
>  > .Internal(inspect(list(x)[rep(1, 2)]))
> @767b080 19 VECSXP g0c2 [NAM(2)] (len=2, tl=0)
>    @7677c18 13 INTSXP g0c1 [NAM(2)] (len=2, tl=0) 1,2
>    @7677c18 13 INTSXP g0c1 [NAM(2)] (len=2, tl=0) 1,2
> 
> but that (creating a list of identical elements) doesn't seem to be a likely 
> real-world scenario and the gain is transient
> 
>  > x = 1:2; y = list(x)[rep(1, 4)]
>  > .Internal(inspect(y))
> @507bef8 19 VECSXP g0c3 [NAM(2)] (len=4, tl=0)
>    @514ff98 13 INTSXP g0c1 [NAM(2)] (len=2, tl=0) 1,2
>    @514ff98 13 INTSXP g0c1 [NAM(2)] (len=2, tl=0) 1,2
>    @514ff98 13 INTSXP g0c1 [NAM(2)] (len=2, tl=0) 1,2
>    @514ff98 13 INTSXP g0c1 [NAM(2)] (len=2, tl=0) 1,2
>  > y[[1]][1] = 2L                  ## everybody copied
>  > .Internal(inspect(y))
> @507bf40 19 VECSXP g0c3 [NAM(1)] (len=4, tl=0)
>    @51502c8 13 INTSXP g0c1 [] (len=2, tl=0) 2,2
>    @51502f8 13 INTSXP g0c1 [] (len=2, tl=0) 1,2
>    @5150328 13 INTSXP g0c1 [] (len=2, tl=0) 1,2
>    @5150358 13 INTSXP g0c1 [] (len=2, tl=0) 1,2
> 
> 
> Probably it is more helpful to think of reducing the number of times an object 
> is _modified_, e.g., representing data as vectors and doing vectorized updates.
> 
> Martin
> 
> >   > mycoef <- list(a=matrix(rnorm(200000), ncol=2000), b=array(rnorm(200000),
> > dim=c(4, 5, 10000)))
> >   > gc()
> >               used   (Mb) gc trigger    (Mb)   max used    (Mb)
> >   Ncells   2650747  141.6    4170209   222.8    4170209   222.8
> >   Vcells 799751724 6101.7 1711485496 13057.6 1711485493 13057.6
> >   > a <- lapply(1:100, function(i) bar(x=mycoef))   # create 100 objects that
> > contain copies
> >   > gc()
> >               used   (Mb) gc trigger    (Mb)   max used    (Mb)
> >   Ncells   2652156  141.7    4170209   222.8    4170209   222.8
> >   Vcells 839752640 6406.9 1711485496 13057.6 1711485493 13057.6
> > # +305 Mb
> >   > b <- lapply(1:100, function(i) foo(x=mycoef))   # same with a reference class
> >   > gc()
> >               used   (Mb) gc trigger    (Mb)   max used    (Mb)
> >   Ncells   2654761  141.8    4170209   222.8    4170209   222.8
> >   Vcells 879756752 6712.1 1711485496 13057.6 1711485493 13057.6
> > # also + 305 Mb
> >   > rm("a", "b")
> >   > gc()
> >               used   (Mb) gc trigger    (Mb)   max used    (Mb)
> >   Ncells   2650660  141.6    4170209   222.8    4170209   222.8
> >   Vcells 799751664 6101.7 1711485496 13057.6 1711485493 13057.6
> > # write to "copy" to see if it uses more memory
> >   > a <- lapply(1:100, function(i) {r <- bar(x=mycoef); r at x$a[5, 10] <- 33; r} )
> >   > gc()
> >               used   (Mb) gc trigger    (Mb)   max used    (Mb)
> >   Ncells   2652174  141.7    4170209   222.8    4170209   222.8
> >   Vcells 839752684 6406.9 1711485496 13057.6 1711485493 13057.6
> > # also + 305 Mb
> >   > rm("a", "b")
> >   Warning message:
> >   In rm("a", "b") : object 'b' not found
> >   > gc()
> >               used   (Mb) gc trigger    (Mb)   max used    (Mb)
> >   Ncells   2650680  141.6    4170209   222.8    4170209   222.8
> >   Vcells 799751684 6101.7 1711485496 13057.6 1711485493 13057.6
> > # now create completely distinct objects
> >   > a <- lapply(1:100, function(i) {acoef <- list(a=matrix(rnorm(200000),
> > ncol=2000), b=array(rnorm(200000), dim=c(4, 5, 10000)))
> > !+                                 bar(x=acoef)})
> >   > gc()
> >               used   (Mb) gc trigger    (Mb)   max used    (Mb)
> >   Ncells   2652191  141.7    4170209   222.8    4170209   222.8
> >   Vcells 839752699 6406.9 1711485496 13057.6 1711485493 13057.6
> > # + 305 Mb
> >
> > Thanks.
> > Ross Boylan
> >
> > P.S. I also tried posting this from a google-managed email account, and have got
> > back two messages like this:
> > Mail Delivery Subsystem mailer-daemon at googlemail.com
> >
> >
> > 5:22 PM (28 minutes ago)
> >
> >
> > to me
> >
> > This is an automatically generated Delivery Status Notification
> >
> > THIS IS A WARNING MESSAGE ONLY.
> >
> > YOU DO NOT NEED TO RESEND YOUR MESSAGE.
> >
> > Delivery to the following recipient has been delayed:
> >
> > r-help at r.project.org <mailto:r-help at r.project.org>
> >
> > Message will be retried for 1 more day(s)
> >
> > Technical details of temporary failure:
> > The recipient server did not accept our requests to connect. Learn more at
> > http://support.google.com/mail/bin/answer.py?answer=7720
> > <http://support.google.com/mail/bin/answer.py?answer=7720>
> > [(0) r.project.org <http://r.project.org>
> > . [206.188.192.100]:25: Connection refused]
> >
> >
> > ______________________________________________
> > R-help at r-project.org mailing list
> > https://stat.ethz.ch/mailman/listinfo/r-help
> > PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> > and provide commented, minimal, self-contained, reproducible code.
> 
>




More information about the R-help mailing list