How to benefit from compiled models

When using an mkin version equal to or greater than 0.9-36 and a C compiler is available, you will see a message that the model is being compiled from autogenerated C code when defining a model using mkinmod. Starting from version, the mkinmod() function checks for presence of a compiler using


In previous versions, it used Sys.which("gcc") for this check.

On Linux, you need to have the essential build tools like make and gcc or clang installed. On Debian based linux distributions, these will be pulled in by installing the build-essential package.

On MacOS, which I do not use personally, I have had reports that a compiler is available by default.

On Windows, you need to install Rtools and have the path to its bin directory in your PATH variable. You do not need to modify the PATH variable when installing Rtools. Instead, I would recommend to put the line

Sys.setenv(PATH = paste("C:/Rtools/bin", Sys.getenv("PATH"), sep=";"))

into your .Rprofile startup file. This is just a text file with some R code that is executed when your R session starts. It has to be named .Rprofile and has to be located in your home directory, which will generally be your Documents folder. You can check the location of the home directory used by R by issuing


Comparison with other solution methods

First, we build a simple degradation model for a parent compound with one metabolite, and we remove zero values from the dataset.

library("mkin", quietly = TRUE)
SFO_SFO <- mkinmod(
  parent = mkinsub("SFO", "m1"),
  m1 = mkinsub("SFO"))
## Temporary DLL for differentials generated and loaded
FOCUS_D <- subset(FOCUS_2006_D, value != 0)

We can compare the performance of the Eigenvalue based solution against the compiled version and the R implementation of the differential equations using the benchmark package. In the output of below code, the warnings about zero being removed from the FOCUS D dataset are suppressed. Since mkin version, an analytical solution is also implemented, which is included in the tests below.

if (require(rbenchmark)) {
  b.1 <- benchmark(
    "deSolve, not compiled" = mkinfit(SFO_SFO, FOCUS_D,
       solution_type = "deSolve",
       use_compiled = FALSE, quiet = TRUE),
    "Eigenvalue based" = mkinfit(SFO_SFO, FOCUS_D,
       solution_type = "eigen", quiet = TRUE),
    "deSolve, compiled" = mkinfit(SFO_SFO, FOCUS_D,
       solution_type = "deSolve", quiet = TRUE),
    "analytical" = mkinfit(SFO_SFO, FOCUS_D,
       solution_type = "analytical",
       use_compiled = FALSE, quiet = TRUE),
    replications = 1, order = "relative",
    columns = c("test", "replications", "relative", "elapsed"))
} else {
  print("R package rbenchmark is not available")
##                    test replications relative elapsed
## 4            analytical            1    1.000   0.181
## 3     deSolve, compiled            1    1.812   0.328
## 2      Eigenvalue based            1    2.088   0.378
## 1 deSolve, not compiled            1   45.923   8.312

We see that using the compiled model is by more than a factor of 10 faster than using deSolve without compiled code.

Model without analytical solution

This evaluation is also taken from the example section of mkinfit. No analytical solution is available for this system, and now Eigenvalue based solution is possible, so only deSolve using with or without compiled code is available.

if (require(rbenchmark)) {
  FOMC_SFO <- mkinmod(
    parent = mkinsub("FOMC", "m1"),
    m1 = mkinsub( "SFO"))

  b.2 <- benchmark(
    "deSolve, not compiled" = mkinfit(FOMC_SFO, FOCUS_D,
                                      use_compiled = FALSE, quiet = TRUE),
    "deSolve, compiled" = mkinfit(FOMC_SFO, FOCUS_D, quiet = TRUE),
    replications = 1, order = "relative",
    columns = c("test", "replications", "relative", "elapsed"))
  factor_FOMC_SFO <- round(b.2["1", "relative"])
} else {
  factor_FOMC_SFO <- NA
  print("R package benchmark is not available")
## Temporary DLL for differentials generated and loaded
##                    test replications relative elapsed
## 2     deSolve, compiled            1    1.000   0.486
## 1 deSolve, not compiled            1   31.597  15.356

Here we get a performance benefit of a factor of 32 using the version of the differential equation model compiled from C code!

This vignette was built with mkin 1.0.3 on

## R version 4.0.3 (2020-10-10)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Debian GNU/Linux bullseye/sid
## CPU model: AMD Ryzen 7 1700 Eight-Core Processor