Simple Memory Profiling in R

Introduction

The profmem() function of the profmem package provides an easy way to profile the memory usage of an R expression. It logs all memory allocations done in R. Profiling memory allocations is helpful when we, for instance, try to understand why a certain piece of R code consumes more memory than expected.

The profmem() function builds upon existing memory profiling features available in R. It logs every memory allocation done by plain R code as well as those done by native code such as C and Fortran. For each entry, it records the size (in bytes) and the name of the functions on the call stack. For example,

> library("profmem")
> options(profmem.threshold = 2000)
> p <- profmem({
+     x <- integer(1000)
+     Y <- matrix(rnorm(n = 10000), nrow = 100)
+ })
> p
Rprofmem memory profiling of:
{
    x <- integer(1000)
    Y <- matrix(rnorm(n = 10000), nrow = 100)
}
Memory allocations (>= 2000 bytes):
       what  bytes               calls
1     alloc   4040           integer()
2     alloc  80040 matrix() -> rnorm()
3     alloc   2544 matrix() -> rnorm()
4     alloc  80040            matrix()
total       166664

From this, we find that 4040 bytes are allocated for integer vector x, which is because each integer value occupies 4 bytes of memory. The additional 40 bytes are due to the internal data structure used for each variable R. The size of this allocation can also be confirmed by the value of object.size(x). We also see that rnorm(), which is called via matrix(), allocates 80040 + 2544 bytes, where the first one reflects the 10000 double values each occupying 8 bytes. The second one reflects some unknown allocation done internally by the native code that rnorm() uses. Finally, the following entry reflects the memory allocation of 80040 bytes done by matrix() itself.

An example where memory profiling can make a difference

Assume we want to set a 100-by-100 matrix with missing values except for element (1,1) that we assign to be zero. This can be done as:

> x <- matrix(nrow = 100, ncol = 100)
> x[1, 1] <- 0
> x[1:3, 1:3]
     [,1] [,2] [,3]
[1,]    0   NA   NA
[2,]   NA   NA   NA
[3,]   NA   NA   NA

This looks fairly innocent, but it turns out that it is very inefficient - both when it comes to memory and speed. The reason is that the default value used by matrix() is NA, which is of type logical. This means that initially x is a logical matrix not a numeric matrix. When we the assign the (1,1) element the value 0, which is a numeric, the matrix first has to be coerced to numeric internally and then the zero is assigned. Profiling the memory will reveal this;

> p <- profmem({
+     x <- matrix(nrow = 100, ncol = 100)
+     x[1, 1] <- 0
+ })
> print(p, expr = FALSE)
Memory allocations (>= 2000 bytes):
       what  bytes      calls
1     alloc  40040   matrix()
2     alloc  80040 <internal>
total       120080

The first entry is for the logical matrix with 10,000 elements (= 4 * 10,000 bytes + small header) that we allocate. The second entry reveals the coercion of this matrix to a numeric matrix (= 8 * 10,000 elements + small header).

To avoid this, we make sure to create a numeric matrix upfront as:

> p <- profmem({
+     x <- matrix(NA_real_, nrow = 100, ncol = 100)
+     x[1, 1] <- 0
+ })
> print(p, expr = FALSE)
Memory allocations (>= 2000 bytes):
       what bytes    calls
1     alloc 80040 matrix()
total       80040

Using the microbenchmark package, we can also quantify the extra overhead in processing time that is introduced due to the logical-to-numeric coercion;

> library("microbenchmark")
> stats <- microbenchmark(bad = {
+     x <- matrix(nrow = 100, ncol = 100)
+     x[1, 1] <- 0
+ }, good = {
+     x <- matrix(NA_real_, nrow = 100, ncol = 100)
+     x[1, 1] <- 0
+ }, times = 100, unit = "ms")
> stats
Unit: milliseconds
 expr   min    lq  mean median    uq   max neval cld
  bad 0.015 0.016 0.047  0.016 0.043 1.457   100   a
 good 0.010 0.011 0.016  0.011 0.016 0.036   100   a

The ineffcient approach is 1.5-2 times slower than the efficient one.

The above illustrates the value of profiling your R code's memory usage and thanks to profmem() we can compare the amount of memory allocated of two alternative implementations. Being able to write memory-efficient R code becomes particularly important when working with large data sets, where an inefficient implementation may even prevent us from performing an analysis because we end up running out of memory. Moreover, each memory allocation will eventually have to be deallocated and in R this is done automatically by the garbage collector, which runs in the background and recovers any blocks of memory that are allocated but no longer in use. Garbage collection takes time and therefore slows down the overall processing in R even further.

What is logged?

The profmem() function uses the utils::Rprofmem() function for logging memory allocation events to a temporary file. The logged events are parsed and returned as an in-memory R object in a format that is convenient to work with. All memory allocations that are done via the native allocVector3() part of R's native API are logged, which means that nearly all memory allocations are logged. Any objects allocated this way are automatically deallocated by R's garbage collector at some point. Garbage collection events are not logged by profmem(). Allocations not logged are those done by non-R native libraries or R packages that use native code Calloc() / Free() for internal objects. Such objects are not handled by the R garbage collector.

Difference between `utils::Rprofmem()` and `utils::Rprof(memory.profiling = TRUE)`

In addition to utils::Rprofmem(), R also provides utils::Rprof(memory.profiling = TRUE). Despite the close similarity of their names, they use completely different approaches for profiling the memory usage. As explained above, the former logs all individual (allocVector3()) memory allocation whereas the latter probes the total memory usage of R at regular time intervals. If memory is allocated and deallocated between two such probing time points, utils::Rprof(memory.profiling = TRUE) will not log that memory whereas utils::Rprofmem() will pick it up. On the other hand, with utils::Rprofmem() it is not possible to quantify the total memory usage at a given time because it only logs allocations and does therefore not reflect deallocations done by the garbage collector.

Requirements

In order for profmem() to work, R must have been built with memory profiling enabled. If not, profmem() will produce an error with an informative message. To manually check whether an R binary was built with this enable or not, do:

> capabilities("profmem")
profmem 
   TRUE

The overhead of running an R installation with memory profiling enabled compared to one without is neglectable / non-measurable.

Volunteers of the R Project provide and distribute pre-built binaries of the R software for all the major operating system via CRAN. It has been confirmed that the R binaries for Windows, macOS (both by CRAN and by the AT&T Research Lab), and for Linux (*) all have been built with memory profiling enabled. (*) For Linux, this has been confirmed for the Debian/Ubuntu distribution but yet not for the other Linux distributions.

In all other cases, to enable memory profiling, which is only needed if capabilities("profmem") returns FALSE, R needs to be configured and built from source using:

$ ./configure --enable-memory-profiling
$ make