Skip to contents

NMsim and speed

Source: vignettes/NMsim-speed.Rmd
NMsim-speed.Rmd

The following points show ways to improve calculation speed with NMsim. The list is somewhat prioritized, most important first.

  • Reduce amount of data written to file Almost always use the table.vars argument. See the example below too.

  • Break the simulation into multiple Nonmem runs

  • Use data.table Depending on the data sets this can make a very large difference. Consider using NMdataConf(as.fun="data.tabe")

  • Use method.execute="nmsim"

  • Parellilize individual runs NMsim does support parellilization (performing a Nonmem run using multiple cores). However, this is rarely worth it for simulations. Most of the time, input/output (reading/writing to disk) is the bottleneck, and splitting the data set into multiple runs will most often reduce run times more.

Most common reasons for NMsim to fail or be slow

If NMsim can find the Nonmem executable and/or PSN, it should work on basically any models where where it makes sense to replace the $ESTIMATION by a $SIMULATION. If NMsim fails or behaves unexpectedly, make sure to read the output from Nonmem in the R console. If NMsim complains it cannot find the output tables, it is likely because Nonmem failed and did not generate them.

There is one thing about the Nonmem model evaluation you have to remember. Nonmem cannot run a model if it uses variables that are unavailable. It may be a covariate in the model that you did not include in your data set. And often the estimation models include variables in output tables that were not used for anything else by Nonmem than being read from the input data set and printed in output tables. In fact, this was exactly why we included a row identifier calle ROW when generating the simulation data set. If we do not do that, we get this error:

Starting NMTRAN
 
 AN ERROR WAS FOUND IN THE CONTROL STATEMENTS.
 
AN ERROR WAS FOUND ON LINE 60 AT THE APPROXIMATE POSITION NOTED:
 $TABLE ROW TVKA TVV2 TVV3 TVCL KA V2 V3 CL Q PRED IPRED Y NOPRINT FILE=NMsim_xgxr021_noname.tab
         X  
 THE CHARACTERS IN ERROR ARE: ROW
  479  THIS ITEM IS NOT LISTED IN MODULE NMPRD4 AND MAY NOT BE DISPLAYED.
cp: cannot stat 'NMsim_xgxr021_noname.tab': No such file or directory
Error in NMscanTables(file, quiet = TRUE, as.fun = "data.table", col.row = col.row,  : 
  NMscanTables: File not found: /home/philip/R/x86_64-pc-linux-gnu-library/4.2/NMsim/examples/nonmem/NMsim/xgxr021_noname/NMsim_xgxr021_noname.tab. Did you copy the lst file but forgot table file?
Results could not be read.

Nonmem gets to writing the $TABLE but cannot find a variable called ROW. Again, we fixed that by including ROW. In many cases, the best way to fix this is to reduce the $TABLE section using the table.vars argument. All we need from the simulation results are population and individual predictions anyway. We could have omitted ROW in the input data set and done something as simple as

simres <- NMsim(file.mod=file.mod,
                data=dat.sim,
                table.vars="PRED IPRED Y")

table.vars can help avoid many of these problems. And if NMsim is slow, this is a potentially a very large low-hanging fruit. In a benchmark example, I reduced a (very large) simulation run time from ~1.5 hours to ~7 minutes this way.