By-Group Aggregation in Parallel

Similar to the row search, by-group aggregation is another perfect use case to demonstrate the power of split-and-conquer with parallelism.

In the example below, it is shown that the homebrew by-group aggregation with foreach pakage, albeit inefficiently coded, is still a lot faster than the summarize() function in Hmisc package.

load('2008.Rdata')

pkgs <- c('rbenchmark', 'doParallel', 'foreach', 'Hmisc')
lapply(pkgs, require, character.only = T)
registerDoParallel(cores = 8)

benchmark(replications = 10,
  summarize = {
    summarize(data[c("Distance", "Month")], data["Month"], colMeans, stat.name = NULL)
  },
  foreach = {
    data2 <- split(data, data$Month)
    test2 <- foreach(i = data2, .combine = rbind) %dopar% (data.frame(Month = unique(i$Month), Distance= mean(i$Distance)))
  }
)

#        test replications elapsed relative user.self sys.self user.child
# 2   foreach           10  19.644     1.00    17.411    1.965      1.528
# 1 summarize           10  30.244     1.54    29.822    0.441      0.000