[R] Naming functions for the purpose of profiling

[Magnus Torfason]

I wanted to send an update on my exploits in trying to get rid of 
<Anonymous> entries in my Rprof.out file (see my prior mail below). 
Basically, I have now found a way to dynamically name a function, and 
even if it is not the cleanest way, I think it will help me in some of 
my profiling needs, and may help others as well. I'm including some 
demonstration code that addresses the issues I wanted to solve:

##########################
# Demonstration of function names in profiling output
a.func <- function()  for (i in 1:1000000) b = 100^100

# Here, the function name ends up in Rprof.out
# (This is the straight-forward example)
Rprof("Rprof-simple-a.out")
a.func()
Rprof(NULL)

# This doesn't work
# (Function shows up as '<Anonymous>' in Rprof.out)
b = list()
b$func = a.func
Rprof("Rprof-simple-b.out")
b$func()
Rprof(NULL)

# By assigning to a variable, the anonymous function gets an
# explicit name (Function shows up as 'c.func' in Rprof.out)
c.func = b$func
Rprof("Rprof-simple-c.out")
c.func()
Rprof(NULL)

# This example is a bit contrived, but works nevertheless
# The anonymous function is assigned to an explicit variable,
# but the variable name can by generated dynamically
# (Function shows up as 'd.func' in Rprof.out, and the stack
# also includes calls to 'eval' as expected)
d = list()
d$func = b$func
Rprof("Rprof-simple-d.out")
desired.func.name = "d.func"
eval(parse(text=paste(
     desired.func.name," = d$func; ",desired.func.name,"()",
     sep="")))
Rprof(NULL)
##########################

And thanks to Jim Holtman who contacted me off-line and gave me some 
helpful advice on profiling in general.

Best,
Magnus


On 1/5/2010 2:58 PM, Magnus Torfason wrote:
> Hi all,
>
> I have some long-running code that I'm trying to profile. I am seeing a
> lot of time spent inside the <Anonymous> function. Of course, this can
> in fact be any of several functions, but I am unable to see how I could
> use the information from Rprof.out to discern which function is taking
> the most time. An example line from my Rprof.out is:
>
> rbernoulli <Anonymous> runOneRound FUN lapply sfLapply doTryCatch
> tryCatchOne tryCatchList tryCatch try eval.with.vis eval.with.vis source
>
> In my case, the <Anonymous> functions seem to be any of several
> work-horse functions, that are part of a list, and different cases are
> dispatched to different functions. I could of course get them all out of
> the list and rewrite the dispatch code, but that does not seem like a
> neat way do address this. So I am wondering if there is any way to
> explicitly set the name of a function in a way that leads to it being
> picked up by the profiler?
>
> The same problem seems to apply to <Anonymous> functions that are
> generated on the fly in an apply call or elsewhere.
>
> Of course, having source files and line numbers included in the
> profiling output would solve this issue, but it seems to me that R
> probably does not have any awareness of where a function was "written
> down" (and of course, even the text of a function can be constructed
> dynamically within a program). So I guess that is not a viable approach.
>
> Thanks in advance for any help on this, and any pointers on the best
> references for advanced profiling issues would be appreciated as well (I
> know of summaryRprof of course, but it can be difficult to get the full
> picture from the summaryRprof output if the calling structure is
> complicated).
>
> Best,
> Magnus