Experiment

• Take range of possible classes number (n_class) as 1:20.
• For every value of n_class generate functions for “if-else” (if_get_true), “switch” (switch_get_true) and “generic+methods” (gen_get_true) approaches. Each of this function will take one argument x. They will check the class of x and perform return(TRUE) action (regardless of class).
• Measure for particular n_class (with package microbenchmark) computation time of all possible *_get_true(x):
  • * can be if, switch and gen.
  • x is constructed independently from * and is a numeric value 1 with class equals to one of class1, …, class{n_class} (total n_class possibilities).
  • Argument times of microbenchmark (actual number of times to evaluate the expression) should be quite big. As preliminary tests showed, computation time differs in microseconds, and big number of actual evaluations is needed to remove the statistical noise. In this analysis it is taken as \(1000000\).
  • The final benchmark of computation time is median of all actual evaluation times (in microseconds).
• Totally there should be \((1 + 2 + … + 20) \cdot 3 = 630\) measurements.

Hypotheses

1. “Generic+methods” benchmark correlates with number of classes. Search of appropriate method in bigger set should take more time.
2. “Generic+methods” benchmark doesn’t correlate with class value. Just a guess.
3. “If-else” benchmark positively correlates with class value. The later class is in “if-else” sequence the more time it should take to get to it.
4. “If-else” and “switch” benchmarks are better than “generic+methods” with small number of classes but the opposite is true with bigger ones. This is just the initial guess which led to this analysis.

Setup

suppressMessages(library(tidyverse))
suppressMessages(library(rlang))
suppressMessages(library(microbenchmark))

set.seed(1105)

ggplot2::theme_set(theme_bw())

Function Generators

Function new_get_true_all() takes number of possible classes n_class and target environment env (by default, it is the environment from which function is called). It creates all necessary functions in env and returns it for cleaner use inside dplyr’s mutate.

# Wrapper for creating function with one argument `x` in environment `env`
new_f <- function(name, body, env) {
  fun <- new_function(alist(x = ), parse_expr(body), env)
  
  assign(x = name, value = fun, envir = env)
}

new_if_get_true <- function(n_class = 1, env = caller_env()) {
  body <- paste0(
    'if (class(x) == "class', seq_len(n_class), '") { return(TRUE) }',
    collapse = " else "
  )
  
  new_f("if_get_true", body, env)
}

new_switch_get_true <- function(n_class = 1, env = caller_env()) {
  body <- paste0(
    "switch(\nclass(x),\n",
    paste0("class", seq_len(n_class), " = return(TRUE)",
           collapse = ",\n"),
    "\n)"
  )
  
  new_f("switch_get_true", body, env)
}

new_gen_get_true <- function(n_class = 1, env = caller_env()) {
  # Create generic
  new_f("gen_get_true", 'UseMethod("gen_get_true")', env)
  
  # Create methods
  method_names <- paste0("gen_get_true.class", seq_len(n_class))
  
  walk(method_names, new_f, body = "return(TRUE)", env = env)
}

new_get_true_all <- function(n_class = 1, env = caller_env()) {
  new_if_get_true(n_class = n_class, env = env)
  new_switch_get_true(n_class = n_class, env = env)
  new_gen_get_true(n_class = n_class, env = env)
  
  env
}

For example, the result of calling new_get_true_all(n_class = 2) from console is creation of the following functions in global environment (here class1 has id 1, class2 - 2 and so on):

new_get_true_all(n_class = 2)
## <environment: R_GlobalEnv>

if_get_true
## function (x) 
## if (class(x) == "class1") {
##     return(TRUE)
## } else if (class(x) == "class2") {
##     return(TRUE)
## }

switch_get_true
## function (x) 
## switch(class(x), class1 = return(TRUE), class2 = return(TRUE))

gen_get_true
## function (x) 
## UseMethod("gen_get_true")

gen_get_true.class1
## function (x) 
## return(TRUE)

gen_get_true.class2
## function (x) 
## return(TRUE)

Benchmark

Function for creating benchmarks for one value of n_class given already created environment env with all functions needed:

bench_funs <- function(n_class = 1, env = caller_env(), times = 1000000) {
  bench <- map(seq_len(n_class), function(class_id) {
    assign("x", structure(1, class = paste0("class", class_id)), envir = env)
    assign("times", times, envir = env)
    
    eval(
      quote(microbenchmark(
        'if' = if_get_true(x),
        'switch' = switch_get_true(x),
        gen = gen_get_true(x),
        times = times
      )),
      envir = env
    ) %>%
      as_tibble() %>%
      group_by(expr) %>%
      # Median computation time in microseconds
      summarise(time = median(time) / 1000) %>%
      mutate(class_id = class_id)
  }) %>%
    bind_rows() %>%
    rename(method = expr)
  
  rm(list = c("x", "times"), envir = env)
  
  bench
}

Computing benchmarks:

# Takes considerable amount of time to run
overhead_bench <- tibble(n_class = 1:20) %>%
  mutate(
    env = rerun(n(), child_env(.GlobalEnv)),
    env = map2(n_class, env, new_get_true_all),
    bench = map2(n_class, env, bench_funs, times = 1000000)
  ) %>%
  select(-env) %>%
  unnest(bench) %>%
  mutate(method = as.character(method)) %>%
  select(n_class, class_id, method, time)

The result has the following structure:

overhead_bench
## # A tibble: 630 x 4
##    n_class class_id method  time
##      <int>    <int>  <chr> <dbl>
##  1       1        1     if 0.529
##  2       1        1 switch 0.381
##  3       1        1    gen 0.820
##  4       2        1     if 0.521
##  5       2        1 switch 0.396
##  6       2        1    gen 0.811
##  7       2        2     if 0.961
##  8       2        2 switch 0.544
##  9       2        2    gen 1.029
## 10       3        1     if 0.554
## # ... with 620 more rows

Plots

Let’s define function for plotting median computation time for every approach based on parameter param (which will be n_class and class_id).

plot_median_time <- function(tbl, param) {
  param_enquo <- enquo(param)
  
  overhead_bench %>%
    mutate(
      Method = case_when(
        method == "gen" ~ "generic+\nmethods",
        method == "if" ~ "if-else",
        method == "switch" ~ "switch"
      )
    ) %>%
    group_by(Method, param = UQ(param_enquo)) %>%
    summarise(median_time = median(time)) %>%
    ungroup() %>%
    ggplot(aes(param, median_time, group = Method, colour = Method)) +
      geom_point() + geom_line() +
      geom_hline(yintercept = 0, colour = "red")
}
overhead_bench %>%
  plot_median_time(n_class) +
    labs(
      title = "Benchmarks for number of classes",
      x = "Number of possible classes",
      y = "Median computation time (microseconds)"
    )

overhead_bench %>%
  plot_median_time(class_id) +
    labs(
      title = "Benchmarks for class id",
      x = 'Class id (number in sequence of "if-else"s)',
      y = "Median computation time (microseconds)"
    )

Correlations

Similarly to plotting, let’s define a function for computing correlation coefficient CI for parameter benchmarks.

extract_cor_ci <- function(cor_test) {
  ci <- round(cor_test$conf.int, digits = 4)
  
  tibble(lower = ci[1], upper = ci[2])
}

compute_median_time_cor_ci <- function(tbl, param) {
  param_enquo <- enquo(param)
  
  tbl %>%
    group_by(method, UQ(param_enquo)) %>%
    summarise(median_time = median(time)) %>%
    summarise(cor_test = list(cor.test(UQ(param_enquo), median_time,
                                       conf.level = 0.95))) %>%
    mutate(cor_ci = map(cor_test, extract_cor_ci)) %>%
    select(-cor_test) %>%
    unnest(cor_ci)
}
compute_median_time_cor_ci(overhead_bench, n_class)
## # A tibble: 3 x 3
##   method   lower  upper
##    <chr>   <dbl>  <dbl>
## 1    gen -0.4407 0.4444
## 2     if  0.9264 0.9886
## 3 switch  0.8297 0.9726
compute_median_time_cor_ci(overhead_bench, class_id)
## # A tibble: 3 x 3
##   method   lower   upper
##    <chr>   <dbl>   <dbl>
## 1    gen -0.8812 -0.4056
## 2     if  0.9912  0.9987
## 3 switch  0.9395  0.9907