The parallel apply family

Parallel Programming in R

Nabeel Imam

Data Science

Weight gain during pregnancy

print(ls_weights)

$AK
  [1] 30 33 50 30 26 99 52 28 24 40 30 99 40
 [14] 20 20 39 25 30 20 33 99 54 30 99 30 38
...
$AL
   [1] 60 32 30 21 30 27 28 32 18 28 30 24 59
  [14] 30 19 30 15 17 18 53 26  5 25 41 50 20
...
$AR
  [1] 31 40 30 99 27 36 30 35 25 19 10 25 40
 [14] 33 37 12 34 99 43 26 28 99 37 30 26  6
...

¹ Natality Data from CDC Births

FORKing with mclapply

boot_mean <- function (weights) {

  est <- rep(0, 1e4)

  for (i in 1:1e4) {
    b <- sample(weights, replace = T)
    est[i] <- mean(b)
  }

  return(est)
}

ls_boot <- mclapply(ls_weights, boot_mean,

                    mc.cores = 4)

Faster by 20 to 40%
FORK cluster generated in the background
- No cluster set up required
- Does not work on Windows

Balance the load with parLapplyLB

Two office workers have a heavy workload, while another worker is idling.

lapply(ls_weights, length)

   AK    AL    AR    AZ    CA    CO    CT   DC    DE    FL    GA    HI    IA  ...
  260  1503   816  1649 14869  1286  1151  520   295  4693  2824   510   891  ...

¹ Designed by Freepik

Balance the load with parLapplyLB

microbenchmark(
  "parLapply" = {
    cl <- makeCluster(4)
    ls <- parLapply(cl, ls_weights, boot_mean)
    stopCluster(cl)
  },
  "parLapplyLB" = {
    cl <- makeCluster(4)
    ls <- parLapplyLB(cl, ls_weights, boot_mean)
    stopCluster(cl)
  },
  times = 10
)

Unit: seconds
       expr      min     mean      max neval
  parLapply 15.77250 16.38548 16.83822    10
parLapplyLB 14.55609 15.49170 17.26866    10

Average (mean) execution time drops from 16.4 seconds to 15.5 seconds (~ 5% reduction)

Multiple arguments to loop over

add <- function (x, y, z) x + y + z


value1 <- c(0.5, 3.2, 5.1, 1.9)
value2 <- c(0.1, 0.5, 0.2, 2.4)
value3 <- 5

For the first iteration, the first elements of value1, value2, and value3 are added. In the second iteration, the second elements of each vector are added, and so on.

clusterMap

cl <- makeCluster(4)


clusterMap(cl, add,

           x = value1,
           y = value2,
           z = value3) # Single value

stopCluster(cl)

print(value3)

[1] 5

[[1]]
[1] 5.6

[[2]]
[1] 8.5

[[3]]
[1] 10.2

[[4]]
[1] 8.4

clusterMap vs. parLapply

clusterMap(cl,         # Cluster
           add,        # Function
           x = value1, # Multiple inputs
           y = value2,
           z = value3) # Could be single values to recycle

parLapply(cl,                # Cluster
          input,             # One input
          fun,               # Function
          arg_name = static) # Exported static variables
                             # supplied to named argument

Weight gained with clusterMap

print(ls_weights)

$AK
  [1] 30 33 50 30 26 99 52 28 24 40 30 99 40
...

print(ls_plur)

$AK
  [1] 1 2 1 1 1 1 1 1 2 1 1 1 1 1 1 1 1 1 1 1
  ...

boot_dist <- function (weights, pluralities) {
  est <- rep(0, 1e4)
  ratio <- weights/pluralities
  for (i in 1:1e4) {
    b <- sample(ratio, replace = T)
    est[i] <- mean(b)
  }
  return(est)
}


cl <- makeCluster(4)

result <- clusterMap(cl, boot_dist,

                     weights = ls_weights,
                     pluralities = ls_plur)

stopCluster(cl)

Total number of births

print(mat_births)

     1   2   3   4   5   6   7   8   9  10  11  12
AK   7  13   8   8   5   9  12  19  16  16   7  16
AL  55  58  79  66  48  63  69  81  74  76  61  82
AR  28  28  41  37  40  40  42  52  40  44  33  38
AZ  69  70  70  71  68  61  73  81  76  74  87  62
CA 603 599 659 601 669 719 675 687 660 698 703 633
CO  50  55  72  56  63  56  56  70  58  48  53  58
CT  51  51  57  58  59  41  51  46  54  56  58  52
DC  25  18  14  40  22  19  24  25  20  26  20  25
...

Row and column operations in parallel

cl <- makeCluster(4)


# parCapply for column operations
parCapply(cl, mat_births, sum)



# parRapply for row operations
parRapply(cl, mat_births, sum)

stopCluster(cl)

Column-wise sum

 1    2    3    4    5    6    7
4258 3966 4437 4345 4371 4482 4614
 8    9   10   11   12
4738 4686 4483 4427 4247

Row-wise sum

AK   AL   AR   AZ   CA   CO   CT
136  812  463  862 7906  695  634
DC   DE   FL   GA   HI   IA   ID
278  151 2519 1512  254  472  237
...

Let's practice!

Parallel Programming in R