On 29/05/2012 09:57, Alon Wasserman wrote:> Hi,
> We've encountered a difference in running time between a straight
function
> call and the same call using do.call when the called function generated an
> error. We've isolated the problem to the following small reproducible
> example:
>
> Consider the following function:
> foo<- function(nr = 2e6, nc=3, use.do.call = FALSE) {
> nn<- paste("V", 1:nc, sep="")
> z<- data.frame(matrix(rnorm(nr*nc), nrow=nr, ncol = nc, dimnames >
list(NULL, nn)))
>
> foo2<- function(x) x[,"V1"] + x[,"V0"]
> if (use.do.call)
> do.call(foo2, list(z))
> else
> foo2(z)
> }
>
> foo2, when called, generates an error because it accesses the V0 column
> which does not exist. When use.do.call==FALSE, foo2 is called directly.
> When use.do.call==TRUE, foo2 is called with the same arguments but using
> do.call. Calling foo() takes about 1 second. Calling
> foo(use.do.call=TRUE)takes about 20 seconds. Does anybody know what
> could explain the difference
> in running time? The difference seems to be related to error handling,
> since try(foo(use.do.call=TRUE)) takes just 1 second.
>
> We used the latest R version (2.15.0) for the test.
>
> Any insight will be appreciated,
Try traceback(max.lines = 10) after each.
With do.call it gives the error equally soon: there is deparsing to be
done in reporting the calls. That is because you passed the object z
and not the symbol z: see the help for do.call.
There are better ways to construct calls: including as here not doing so
but presumably not in the real problem. See e.g. ?call and ?substitute
I don't really see why you did not post this on R-help: it is entirely
about R programming.
> Thanks,
> Alon
--
Brian D. Ripley, ripley at stats.ox.ac.uk
Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/
University of Oxford, Tel: +44 1865 272861 (self)
1 South Parks Road, +44 1865 272866 (PA)
Oxford OX1 3TG, UK Fax: +44 1865 272595