Is S4 method dispatch slow?

Question

My S4 class has a method that is called many times. I noticed that the execution time is much slower than it would be if a similar function was called independently. So I ad

Answer 1

The cost is in method look-up, which starts from scratch in each iteration of your timing. This can be short-circuited by figuring out method dispatch once

METHOD <- selectMethod(method.foo, class(st))
for (i in seq(iters)) METHOD(st)

This (better method look-up) would be a very interesting and worth-while project; there are valuable lessons learned in other dynamic languages, e.g., inline caching mentioned on Wikipedia's dynamic dispatch page.

I wonder if the reason that you're making many method calls is because of incomplete vectorization of your data representation and methods?

Answer 2

This doesn't help you directly with your problem, but it's much easier to benchmark this sort of stuff with the microbenchmark package:

f <- function(x) NULL

s3 <- function(x) UseMethod("s3")
s3.integer <- function(x) NULL

A <- setClass("A", representation(a = "list"))
setGeneric("s4", function(x) standardGeneric("s4"))
setMethod(s4, "A", function(x) NULL)

B <- setRefClass("B")
B$methods(r5 = function(x) NULL)

a <- A()
b <- B$new()

library(microbenchmark)
options(digits = 3)
microbenchmark(
  bare = NULL,
  fun = f(),
  s3 = s3(1L),
  s4 = s4(a),
  r5 = b$r5()
)
# Unit: nanoseconds
#  expr   min    lq median    uq   max neval
#  bare    13    20     22    29    36   100
#   fun   171   236    270   310   805   100
#    s3  2025  2478   2651  2869  8603   100
#    s4 10017 11029  11528 11905 36149   100
#    r5  9080 10003  10390 10804 61864   100

On my computer, the bare call takes about 20 ns. Wrapping it in a function adds about an extra 200 ns - this is the cost of creating the environment where the function execution happens. S3 method dispatch adds around 3 µs and S4/ref classes around 12 µs.