xrprof
(formerly rtrace
) is an external sampling profiler for R on Linux and
Windows.
Many R users will be familiar with using the built-in sampling profiler
Rprof()
to generate data on what their code is doing, and there are several
excellent tools to facilitate understanding these samples (or serve as a
front-end), including the profvis
package.
However, the reach of Rprof()
and related tools is limited: the profiler is
"internal", in the sense that it must be manually switched on to work, either
during interactive work (for example, to profile an individual function), or
perhaps by modifying the script to include Rprof()
calls before running it
again.
In contrast, xrprof
can be used to profile code that is already running:
$ Rscript myscript.R &
# sudo may be required.
$ xrprof -p <PID> -F 50 > Rprof.out
External sampling profilers have proven extremely useful for diagnosing and
fixing performance issues (or other bugs) in production environments. This
project joins a large list similar tools for other languages, such as perf
(the Linux system profiler), jstack
(for Java), rbspy
(for Ruby), Pyflame
(for Python), VSPerfCmd
for C#/.NET, and many others.
xrprof
depends on libelf and libunwind, so you must have their headers to
compile the program. For example, on Debian-based systems (including Ubuntu),
you can install these with
$ sudo apt-get install libelf-dev libunwind-dev libcap2-bin
A simple Makefile
is provided. Build the binary with
$ git clone https://github.com/atheriel/xrprof
$ cd xrprof
$ make
To install the profiler to your system, use
$ sudo make install
This will install the binary to /usr/local/bin
and use setcap
to mark it for
use without sudo
. The install
target supports prefix
and DESTDIR
.
You must have a build environment set up. For R users, the best option is to use R's own Rtools for Windows (which is also used to install packages from source). You can then launch "Rtools MinGW 64-bit" from the Start Menu and navigate to the source directory; then run
$ git clone https://github.com/atheriel/xrprof
$ cd xrprof
$ make -f Makefile.win
The resulting xrprof.exe
program can be run from cmd.exe
or PowerShell.
The profiler has a simple interface:
Usage: xrprof [-F <freq>] [-d <duration>] -p <pid>
The Rprof.out
format is written to standard output and errors or other
messages are written to standard error.
On Windows, R's process ID (PID) can be looked up in Task Manager.
Along with the sampling profiler itself, there is also a stackcollapse-Rprof.R
script in tools/
that converts the Rprof.out
format to one that can be
understood by Brendan Gregg's FlameGraph
tool. You can use this to produce graphs like the one below:
$ stackcollapse-Rprof.R Rprof.out | flamegraph.pl > Rprof.svg
A public Docker image is available at atheriel/xrprof
. Since xrprof
reads
the memory of other running programs, it must be run as a privileged container
in the host PID namespace. For example:
$ docker run --privileged --pid=host -it atheriel/xrprof -p <PID>
Much like other sampling profilers, the program uses Linux's ptrace
system
calls to attach to running R processes and a mix of ptrace
and
process_vm_readv
to read the memory contents of that process, following
pointers along the way.
The R-specific aspect of this is to locate and decode the R_GlobalContext
structure inside of the R interpreter that stores information on the currently
executing R code.
In order to defeat address space randomization, xrprof
will search through the
ELF files loaded into memory (at /proc/<pid>/maps
) for the symbols required,
either in the executable itself or in libR.so
(if it appears R has been
compiled to use it).
xrprof
is mount-namespace-aware, so it supports profiling R processes running
inside Docker containers.
On Windows, xrprof
makes use of APIs like ReadProcessMemory()
,
NtSuspendProcess()
, and SymFromName()
to achieve the analogous result.
The project was inspired by Julia Evan's blog posts on writing
rbspy
and later by my discovery of Evan Klitzke's
work (and writing) on Pyflame.
This project contains portions of the source code of R itself, which is copyright the R Core Developers and licensed under the GPLv2.
The remaining code is copyright its authors and also available under the same license, GPLv2.