Process optimisation of code within dopar

问题

I am trying to optimize my code to run glms multiple times, and I would like to leverage parallelization, either with foreach or some other more efficient way.

As you can see; the for loop takes about 800 secs to run 270000 glms; while foreach with dopar unintuitively takes for ever (It either crashes or I force it to stop after a couple of hours).

Thanks for your help.

Jinesh

library(data.table)
library(parallel)
library(doParallel)
library(foreach)
scen_bin <- expand.grid(n = c(10, 20, 30), rate1 = c(0.1, 0.2, 0.3),
  rate2 = c(0.5, 0.6, 0.9))

rep <- 10000
scen_sims <- rbindlist(replicate(rep, scen_bin, simplify = FALSE),
  idcol = TRUE)
scen_sims[, `:=`(glm, list(c(1L, 2L)))]

for (i in 1:270000) {
  set(scen_sims, i, 8L, list(glm(formula = c(rbinom(scen_sims$drug[i], 1L, scen_sims$Treatment_Rates[i]),
    rbinom(scen_sims$control[i], 1L, scen_sims$Comparator_Rates[i])) ~ factor(c(rep("Trt",
    scen_sims$drug[i]), rep("Cont", scen_sims$control[i]))), family = "binomial")))
}

split_scen_sims <- split(scen_sims, seq(1, 270000, length.out = 1000))


jh <- foreach(x = 1:1000, .packages = c("data.table")) %dopar% {
  jh <- split_scen_sims[[x]]
  for (i in 1:270000) {
    set(jh, i, 8L, list(glm(formula = c(rbinom(jh$n[i], 1L, jh$rate1[i]), rbinom(jh$n[i],
      1L, jh$rate1[i])) ~ factor(c(rep("Trt", jh$n[i]), rep("Cont", jh$n[i]))),
      family = "binomial")))
  }
  return(jh)
}

回答1:

The first thing to note is that using the extract function $ within a loop makes this perform poorly. It would be better to 1) make a functions and then 2) use a regular data.table call.

fx_make_glm = function(drug, treat_rate, control, Comparator_Rates){
  glm(formula = c(rbinom(drug, 1L, treat_rate),
                  rbinom(control, 1L, Comparator_Rates)) ~
        factor(c(rep("Trt", drug), rep("Cont", control))), 
      family = "binomial")
}

This will greatly simplify the rest - I will use Map which will loop through each element of the variables of interest:

scen_sims[, glm := list(Map(fx_make_glm, n, rate1, n, rate2))]

Unfortunately, that still didn't provide as much performance as ideal :(

Unit: seconds
     expr  min   lq mean median   uq  max neval
  OP_loop 3.01 3.21 3.21   3.22 3.26 3.36     5
 map_call 2.64 2.89 2.90   2.92 2.96 3.08     5

My parallel package of choice is future.apply - just put future_ in front of your *apply series and you have a parallel evaluation:

library(future.apply)
plan(multiprocess)
system.time({
  scen_sims[, glm := list(future_Map(fx_make_glm, n, rate1, n, rate2))]
})

   user  system elapsed 
   1.22    0.13    3.22 

## truncated the microbenchmark call

Unit: seconds
            expr  min   lq mean median   uq  max neval
         OP_loop 2.93 2.98 3.08   3.00 3.18 3.32     5
        map_call 2.65 2.70 2.94   2.89 3.18 3.25     5
 future_map_call 2.84 3.24 3.37   3.43 3.49 3.85     5

I am on Windows with 2 cores / 4 threads. If I were on Linux, I would try plan(multicore) to see if forking processes were more productive.

Data generation:

library(data.table)
## generate data
scen_bin <- expand.grid(n = c(10, 20, 30), rate1 = c(0.1, 0.2, 0.3),
                        rate2 = c(0.5, 0.6, 0.9))

rep <- 50L
scen_sims <- rbindlist(replicate(rep, scen_bin, simplify = FALSE),
                       idcol = TRUE)
scen_sims[, `:=`(glm, list(c(1L, 2L)))]

来源：https://stackoverflow.com/questions/60498786/process-optimisation-of-code-within-dopar

标签

foreach

data.table

doparallel