Chapter 13 Tidying and profiling R code

R code which is worth preserving in a package and perhaps making available for others to use is worth documenting, tidying up and perhaps optimizing. The last two of these activities are the subject of this chapter.


3.1 Tidying R code

R treats function code loaded from packages and code entered by users differently. By default code entered by users has the source code stored internally, and when the function is listed, the original source is reproduced. Loading code from a package (by default) discards the source code, and the function listing is re-created from the parse tree of the function.

Normally keeping the source code is a good idea, and in particular it avoids comments being removed from the source. However, we can make use of the ability to re-create a function listing from its parse tree to produce a tidy version of the function, for example with consistent indentation and spaces around operators. If the original source does not follow the standard format this tidied version can be much easier to read.

We can subvert the keeping of source in two ways.

  1. The option keep.source can be set to FALSE before the code is loaded into R.
  2. The stored source code can be removed by calling the removeSource() function, for example by
    myfun <- removeSource(myfun)

In each case if we then list the function we will get the standard layout.

Suppose we have a file of functions myfuns.R that we want to tidy up. Create a file tidy.R containing

source("myfuns.R", keep.source = FALSE)
dump(ls(all = TRUE), file = "new.myfuns.R")

and run R with this as the source file, for example by R –vanilla < tidy.R or by pasting into an R session. Then the file new.myfuns.R will contain the functions in alphabetical order in the standard layout. Warning: comments in your functions will be lost.

The standard format provides a good starting point for further tidying. Although the deparsing cannot do so, we recommend the consistent use of the preferred assignment operator ‘<-’ (rather than ‘=’) for assignment. Many package authors use a version of Emacs (on a Unix-alike or Windows) to edit R code, using the ESS[S] mode of the ESS Emacs package. See R coding standards in R Internals for style options within the ESS[S] mode recommended for the source code of R itself.


3.2 Profiling R code for speed

It is possible to profile R code on Windows and most101 Unix-alike versions of R.

The command Rprof is used to control profiling, and its help page can be consulted for full details. Profiling works by recording at fixed intervals102 (by default every 20 msecs) which line in which R function is being used, and recording the results in a file (default Rprof.out in the working directory). Then the function summaryRprof or the command-line utility R CMD Rprof Rprof.out can be used to summarize the activity.

As an example, consider the following code (from Venables & Ripley, 2002, pp. 225–6).

library(MASS); library(boot)
storm.fm <- nls(Time ~ b*Viscosity/(Wt - c), stormer,
                start = c(b=30.401, c=2.2183))
st <- cbind(stormer, fit=fitted(storm.fm))
storm.bf <- function(rs, i) {
    st$Time <-  st$fit + rs[i]
    tmp <- nls(Time ~ (b * Viscosity)/(Wt - c), st,
               start = coef(storm.fm))
    tmp$m$getAllPars()
}
rs <- scale(resid(storm.fm), scale = FALSE) # remove the mean
Rprof("boot.out")
storm.boot <- boot(rs, storm.bf, R = 4999) # slow enough to profile
Rprof(NULL)

Having run this we can summarize the results by

R CMD Rprof boot.out

Each sample represents 0.02 seconds.
Total run time: 22.52 seconds.

Total seconds: time spent in function and callees.
Self seconds: time spent in function alone.
   %       total       %        self
 total    seconds     self    seconds    name
 100.0     25.22       0.2      0.04     "boot"
  99.8     25.18       0.6      0.16     "statistic"
  96.3     24.30       4.0      1.02     "nls"
  33.9      8.56       2.2      0.56     "<Anonymous>"
  32.4      8.18       1.4      0.36     "eval"
  31.8      8.02       1.4      0.34     ".Call"
  28.6      7.22       0.0      0.00     "eval.parent"
  28.5      7.18       0.3      0.08     "model.frame"
  28.1      7.10       3.5      0.88     "model.frame.default"
  17.4      4.38       0.7      0.18     "sapply"
  15.0      3.78       3.2      0.80     "nlsModel"
  12.5      3.16       1.8      0.46     "lapply"
  12.3      3.10       2.7      0.68     "assign"
 ...
   %        self        %      total
  self    seconds     total   seconds    name
   5.7      1.44       7.5      1.88     "inherits"
   4.0      1.02      96.3     24.30     "nls"
   3.6      0.92       3.6      0.92     "$"
   3.5      0.88      28.1      7.10     "model.frame.default"
   3.2      0.80      15.0      3.78     "nlsModel"
   2.8      0.70       9.8      2.46     "qr.coef"
   2.7      0.68      12.3      3.10     "assign"
   2.5      0.64       2.5      0.64     ".Fortran"
   2.5      0.62       7.1      1.80     "qr.default"
   2.2      0.56      33.9      8.56     "<Anonymous>"
   2.1      0.54       5.9      1.48     "unlist"
   2.1      0.52       7.9      2.00     "FUN"
  ...

This often produces surprising results and can be used to identify bottlenecks or pieces of R code that could benefit from being replaced by compiled code.

Two warnings: profiling does impose a small performance penalty, and the output files can be very large if long runs are profiled at the default sampling interval.

Profiling short runs can sometimes give misleading results. R from time to time performs garbage collection to reclaim unused memory, and this takes an appreciable amount of time which profiling will charge to whichever function happens to provoke it. It may be useful to compare profiling code immediately after a call to gc() with a profiling run without a preceding call to gc.

More detailed analysis of the output can be achieved by the tools in the CRAN packages proftools and profr: in particular these allow call graphs to be studied.


3.3 Profiling R code for memory use

Measuring memory use in R code is useful either when the code takes more memory than is conveniently available or when memory allocation and copying of objects is responsible for slow code. There are three ways to profile memory use over time in R code. All three require R to have been compiled with –enable-memory-profiling, which is not the default, but is currently used for the macOS and Windows binary distributions. All can be misleading, for different reasons.

In understanding the memory profiles it is useful to know a little more about R’s memory allocation. Looking at the results of gc() shows a division of memory into Vcells used to store the contents of vectors and Ncells used to store everything else, including all the administrative overhead for vectors such as type and length information. In fact the vector contents are divided into two pools. Memory for small vectors (by default 128 bytes or less) is obtained in large chunks and then parcelled out by R; memory for larger vectors is obtained directly from the operating system.

Some memory allocation is obvious in interpreted code, for example,

y <- x + 1

allocates memory for a new vector y. Other memory allocation is less obvious and occurs because R is forced to make good on its promise of ‘call-by-value’ argument passing. When an argument is passed to a function it is not immediately copied. Copying occurs (if necessary) only when the argument is modified. This can lead to surprising memory use. For example, in the ‘survey’ package we have

print.svycoxph <- function (x, ...)
{
    print(x$survey.design, varnames = FALSE, design.summaries = FALSE, ...)
    x$call <- x$printcall
    NextMethod()
}

It may not be obvious that the assignment to x$call will cause the entire object x to be copied. This copying to preserve the call-by-value illusion is usually done by the internal C function duplicate.

The main reason that memory-use profiling is difficult is garbage collection. Memory is allocated at well-defined times in an R program, but is freed whenever the garbage collector happens to run.


3.3.1 Memory statistics from Rprof

The sampling profiler Rprof described in the previous section can be given the option memory.profiling=TRUE. It then writes out the total R memory allocation in small vectors, large vectors, and cons cells or nodes at each sampling interval. It also writes out the number of calls to the internal function duplicate, which is called to copy R objects. summaryRprof provides summaries of this information. The main reason that this can be misleading is that the memory use is attributed to the function running at the end of the sampling interval. A second reason is that garbage collection can make the amount of memory in use decrease, so a function appears to use little memory. Running under gctorture helps with both problems: it slows down the code to effectively increase the sampling frequency and it makes each garbage collection release a smaller amount of memory. Changing the memory limits with mem.limits() may also be useful, to see how the code would run under different memory conditions.


3.3.2 Tracking memory allocations

The second method of memory profiling uses a memory-allocation profiler, Rprofmem(), which writes out a stack trace to an output file every time a large vector is allocated (with a user-specified threshold for ‘large’) or a new page of memory is allocated for the R heap. Summary functions for this output are still being designed.

Running the example from the previous section with

> Rprofmem("boot.memprof",threshold=1000)
> storm.boot <- boot(rs, storm.bf, R = 4999)
> Rprofmem(NULL)

shows that apart from some initial and final work in boot there are no vector allocations over 1000 bytes.


3.3.3 Tracing copies of an object

The third method of memory profiling involves tracing copies made of a specific (presumably large) R object. Calling tracemem on an object marks it so that a message is printed to standard output when the object is copied via duplicate or coercion to another type, or when a new object of the same size is created in arithmetic operations. The main reason that this can be misleading is that copying of subsets or components of an object is not tracked. It may be helpful to use tracemem on these components.

In the example above we can run tracemem on the data frame st

> tracemem(st)
[1] "<0x9abd5e0>"
> storm.boot <- boot(rs, storm.bf, R = 4)
memtrace[0x9abd5e0->0x92a6d08]: statistic boot
memtrace[0x92a6d08->0x92a6d80]: $<-.data.frame $<- statistic boot
memtrace[0x92a6d80->0x92a6df8]: $<-.data.frame $<- statistic boot
memtrace[0x9abd5e0->0x9271318]: statistic boot
memtrace[0x9271318->0x9271390]: $<-.data.frame $<- statistic boot
memtrace[0x9271390->0x9271408]: $<-.data.frame $<- statistic boot
memtrace[0x9abd5e0->0x914f558]: statistic boot
memtrace[0x914f558->0x914f5f8]: $<-.data.frame $<- statistic boot
memtrace[0x914f5f8->0x914f670]: $<-.data.frame $<- statistic boot
memtrace[0x9abd5e0->0x972cbf0]: statistic boot
memtrace[0x972cbf0->0x972cc68]: $<-.data.frame $<- statistic boot
memtrace[0x972cc68->0x972cd08]: $<-.data.frame $<- statistic boot
memtrace[0x9abd5e0->0x98ead98]: statistic boot
memtrace[0x98ead98->0x98eae10]: $<-.data.frame $<- statistic boot
memtrace[0x98eae10->0x98eae88]: $<-.data.frame $<- statistic boot

The object is duplicated fifteen times, three times for each of the R+1 calls to storm.bf. This is surprising, since none of the duplications happen inside nls. Stepping through storm.bf in the debugger shows that all three happen in the line

st$Time <- st$fit + rs[i]

Data frames are slower than matrices and this is an example of why. Using tracemem(st$Viscosity) does not reveal any additional copying.


3.4 Profiling compiled code

Profiling compiled code is highly system-specific, but this section contains some hints gleaned from various R users. Some methods need to be different for a compiled executable and for dynamic/shared libraries/objects as used by R packages. We know of no good way to profile DLLs on Windows.


3.4.1 Linux

Options include using sprof for a shared object, and oprofile (see http://oprofile.sourceforge.net/) and perf (see https://perf.wiki.kernel.org/index.php/Tutorial) for any executable or shared object.

3.4.1.1 sprof

You can select shared objects to be profiled with sprof by setting the environment variable LD_PROFILE. For example

% setenv LD_PROFILE /path/to/R_HOME/library/stats/libs/stats.so
R
... run the boot example
% sprof /path/to/R_HOME/library/stats/libs/stats.so \
  /var/tmp/path/to/R_HOME/library/stats/libs/stats.so.profile

Flat profile:

Each sample counts as 0.01 seconds.
  %   cumulative   self              self     total
 time   seconds   seconds    calls  us/call  us/call  name
 76.19      0.32     0.32        0     0.00           numeric_deriv
 16.67      0.39     0.07        0     0.00           nls_iter
  7.14      0.42     0.03        0     0.00           getListElement

rm /var/tmp/path/to/R_HOME/library/stats/libs/stats.so.profile
... to clean up ...

It is possible that root access is needed to create the directories used for the profile data.

3.4.1.2 oprofile and operf

The oprofile project has two modes of operation. In what is now called ‘legacy’ mode, it is uses a daemon to collect information on a process (see below). Since version 0.9.8 (August 2012), the preferred mode is to use operf, so we discuss that first. The modes differ in how the profiling data is collected: it is analysed by tools such as opreport and oppannote in both.

Here is an example on x86_64 Linux using R 3.0.2. File pvec.R contains the part of the examples from pvec in package parallel:

library(parallel)
N <- 1e6
dates <- sprintf('%04d-%02d-%02d', as.integer(2000+rnorm(N)),
                 as.integer(runif(N, 1, 12)), as.integer(runif(N, 1, 28)))
system.time(a <- as.POSIXct(dates, format = "%Y-%m-%d"))

with timings from the final step

   user  system elapsed
  0.371   0.237   0.612

R-level profiling by Rprof shows

                     self.time self.pct total.time total.pct
"strptime"                1.70    41.06       1.70     41.06
"as.POSIXct.POSIXlt"      1.40    33.82       1.42     34.30
"sprintf"                 0.74    17.87       0.98     23.67
...

so the conversion from character to POSIXlt takes most of the time.

This can be run under operf and analysed by

operf R -f pvec.R
opreport
opreport -l /path/to/R_HOME/bin/exec/R
opannotate --source /path/to/R_HOME/bin/exec/R
## And for the system time
opreport -l /lib64/libc.so.6

The first report shows where (which library etc) the time was spent:

CPU_CLK_UNHALT...|
  samples|      %|
------------------
   166761 99.9161 Rdev
        CPU_CLK_UNHALT...|
          samples|      %|
        ------------------
            70586 42.3276 no-vmlinux
            56963 34.1585 libc-2.16.so
            36922 22.1407 R
             1584  0.9499 stats.so
              624  0.3742 libm-2.16.so
...

The rest of the output is voluminous, and only extracts are shown below.

Most of the time within R is spent in

samples  %        image name symbol name
10397    28.5123  R           R_gc_internal
5683     15.5848  R           do_sprintf
3036      8.3258  R           do_asPOSIXct
2427      6.6557  R           do_strptime
2421      6.6392  R           Rf_mkCharLenCE
1480      4.0587  R           w_strptime_internal
1202      3.2963  R           Rf_qnorm5
1165      3.1948  R           unif_rand
675       1.8511  R           mktime0
617       1.6920  R           makelt
617       1.6920  R           validate_tm
584       1.6015  R           day_of_the_week
...

opannotate shows that 31% of the time in R is spent in memory.c, 21% in datetime.c and 7% in Rstrptime.h. The analysis for libc showed that calls to wcsftime dominated, so those calls were cached for R 3.0.3: the time spent in no-vmlinux (the kernel) was reduced dramatically.

On platforms which support it, call graphs can be produced by opcontrol –callgraph if collected via operf –callgraph.

The profiling data is by default stored in sub-directory oprofile_data of the current directory, which can be removed at the end of the session.

Another example, from sm version 2.2-5.4. The example for sm.variogram took a long time:

system.time(example(sm.variogram))
...
   user  system elapsed
  5.543   3.202   8.785

including a lot of system time. Profiling just the slow part, the second plot, showed

  samples|      %|
------------------
   381845 99.9885 R
        CPU_CLK_UNHALT...|
          samples|      %|
        ------------------
           187484 49.0995 sm.so
           169627 44.4230 no-vmlinux
            12636  3.3092 libgfortran.so.3.0.0
             6455  1.6905 R

so the system time was almost all in the Linux kernel. It is possible to dig deeper if you have a matching uncompressed kernel with debug symbols to specify via –vmlinux: we did not.

In ‘legacy’ mode oprofile works by running a daemon which collects information. The daemon must be started as root, e.g.

% su
% opcontrol --no-vmlinux
% (optional, some platforms) opcontrol --callgraph=5
% opcontrol --start
% exit

Then as a user

% R
... run the boot example
% opcontrol --dump
% opreport -l /path/to/R_HOME/library/stats/libs/stats.so
...
samples  %        symbol name
1623     75.5939  anonymous symbol from section .plt
349      16.2552  numeric_deriv
113       5.2632  nls_iter
62        2.8878  getListElement
% opreport -l /path/to/R_HOME/bin/exec/R
...
samples  %        symbol name
76052    11.9912  Rf_eval
54670     8.6198  Rf_findVarInFrame3
37814     5.9622  Rf_allocVector
31489     4.9649  Rf_duplicate
28221     4.4496  Rf_protect
26485     4.1759  Rf_cons
23650     3.7289  Rf_matchArgs
21088     3.3250  Rf_findFun
19995     3.1526  findVarLocInFrame
14871     2.3447  Rf_evalList
13794     2.1749  R_Newhashpjw
13522     2.1320  R_gc_internal
...

Shutting down the profiler and clearing the records needs to be done as root.


3.4.2 Solaris

On 64-bit (only) Solaris, the standard profiling tool gprof collects information from shared objects compiled with -pg.


3.4.3 macOS

Developers have recommended sample (or Sampler.app, which is a GUI version), Shark (in version of Xcode up to those for Snow Leopard), and Instruments (part of Xcode, see https://developer.apple.com/library/mac/#documentation/DeveloperTools/Conceptual/InstrumentsUserGuide/Introduction/Introduction.html).