big_parallelize {bigstatsr}

Split-parApply-Combine

Description

A Split-Apply-Combine strategy to parallelize the evaluation of a function.

Usage

big_parallelize(
  X,
  p.FUN,
  p.combine = NULL,
  ind = cols_along(X),
  ncores = nb_cores(),
  ...
)

Arguments

X

An object of class FBM.

p.FUN

The function to be applied to each subset matrix. It must take a Filebacked Big Matrix as first argument and ind, a vector of indices, which are used to split the data. For example, if you want to apply a function to X[ind.row, ind.col], you may use X[ind.row, ind.col[ind]] in a.FUN.

p.combine

Function to combine the results with do.call. This function should accept multiple arguments (...). For example, you can use c, cbind, rbind. This package also provides function plus to add multiple arguments together. The default is NULL, in which case the results are not combined and are returned as a list, each element being the result of a block.

ind

Initial vector of subsetting indices. Default is the vector of all column indices.

ncores

Number of cores used. Default doesn't use parallelism. You may use nb_cores.

...

Extra arguments to be passed to p.FUN.

Details

This function splits indices in parts, then apply a given function to each part and finally combine the results.

Value

Return a list of ncores elements, each element being the result of one of the cores, computed on a block. The elements of this list are then combined with do.call(p.combine, .) if p.combined is given.

See Also

big_apply bigparallelr::split_parapply

Examples

## Not run:  # CRAN is super slow when parallelism.
  X <- big_attachExtdata()

  ### Computation on all the matrix
  true <- big_colstats(X)

  big_colstats_sub <- function(X, ind) {
    big_colstats(X, ind.col = ind)
  }
  # 1. the computation is split along all the columns
  # 2. for each part the computation is done, using `big_colstats`
  # 3. the results (data.frames) are combined via `rbind`.
  test <- big_parallelize(X, p.FUN = big_colstats_sub,
                          p.combine = 'rbind', ncores = 2)
  all.equal(test, true)

  ### Computation on a part of the matrix
  n <- nrow(X)
  m <- ncol(X)
  rows <- sort(sample(n, n/2)) # sort to provide some locality in accesses
  cols <- sort(sample(m, m/2)) # idem

  true2 <- big_colstats(X, ind.row = rows, ind.col = cols)

  big_colstats_sub2 <- function(X, ind, rows, cols) {
    big_colstats(X, ind.row = rows, ind.col = cols[ind])
  }
  # This doesn't work because, by default, the computation is spread
  # along all columns. We must explictly specify the `ind` parameter.
  tryCatch(big_parallelize(X, p.FUN = big_colstats_sub2,
                           p.combine = 'rbind', ncores = 2,
                           rows = rows, cols = cols),
           error = function(e) message(e))

  # This now works, using `ind = seq_along(cols)`.
  test2 <- big_parallelize(X, p.FUN = big_colstats_sub2,
                           p.combine = 'rbind', ncores = 2,
                           ind = seq_along(cols),
                           rows = rows, cols = cols)
  all.equal(test2, true2)


## End(Not run)