R/speed_analysis.R

In andreas50/uts: Unevenly-spaced Time Series

#############################################
# Speed analysis of R vs. C implementations #
#############################################

# Hardware: i7-2600, 32GB RAM
# Software: Windows 7 Pro 64bit, R 3.5.1, gcc-4.9.3

### num_leq_sorted (8/2018)
# Conclusion: for long input vectors, the C implementation is around 14 times faster
if (0) {
  set.seed(1)
  a <- sort(runif(1e4))
  b <- sort(runif(1e4))
  
  # R vs. C: 2.27s vs. 0.22s
  system.time(for (j in 1:500) num_leq_sorted_R(a, b))
  system.time(for (j in 1:500) num_leq_sorted(a, b))
  
  # Profile C implementation
  # -) ~27% of time spent in C code, rest in argument checking
  Rprof(interval=0.01)
  for (j in 1:1e4) num_leq_sorted(a, b)
  Rprof(NULL)
  summaryRprof()
}


### sorted_union (8/2018)
# Conclusion: for long input vectors, the C implementation is only 2.5 times faster
# In addition, the R implementation a) does less error checking, b) has no numerical tolerance support
if (0) {
  set.seed(1)
  a <- sort(runif(1e4))
  b <- sort(runif(1e4))
  
  # R vs. C: 4.09s vs. 0.99s
  system.time(for (j in 1:2000) sorted_union_R(a, b))
  system.time(for (j in 1:2000) sorted_union(a, b))
  
  # Profile C implementation
  # -) ~35% of time spent in C code, rest in argument checking
  Rprof(interval=0.01)
  for (j in 1:1e4) sorted_union(a, b)
  Rprof(NULL)
  summaryRprof()
  
  # Profile R implementation
  # -) ~70% of time spent in sort()
  Rprof(interval=0.01)
  for (j in 1:5000) sorted_union_R(a, b)
  Rprof(NULL)
  summaryRprof()
}