{smallfactor} update - store factors in bytes and bits rather than integers

smallfactor - store factors in bytes and bits rather than an integer

R-CMD-check

smallfactor is an experiment to see what trade-offs there might be for storing a factor in bytes or bits instead of an integer.

An R integer is 4-bytes, meaning it could hold a factor with two billion levels, which seems like overkill.

bytefactor targets factors with up to 256 levels.

bitfactor will choose the fewest bits to store a factor for factors with 2 up to 32768 (2^15) levels.

What’s in the box

  • bytefactor() store factor values in raw bytes rather than an integer.
  • bitfactor() store factor values as bits within integers - with multiple factors stored in a single integer value. E.g. a factor with 4 levels could be stored as just 2 bits of information, meaning 16 values could be stored in a single R integer (16 * 2 bits = 32 bits = 4 bytes)

Limitations

It seems possible that you could write more methods to make these smaller factors behave similar to regular R factor objects, but this package is not attempting this (yet).

Internally, there’s currently a lot of transferring back-and-forth between these small factors and the standard R factor in order to make use of the printing and subsetting capabilities of the R factor implementation. Much of this back-and-forth could be avoided if effort was expended to do so.

Note: bitfactor uses only 31 bits of a 32-bit integer in order to avoid issues around NA_integer_ representation. This means, for example, that an integer can only hold 15 x 2-bit values. In practice the user is never expected to notice this or care about it.

Installation

You can install from GitHub with:

# install.package('remotes')
remotes::install_github('coolbutuseless/smallfactor')

bytefactor

small <- bytefactor(c('a', 'b', 'c', 'a', 'd'))
small
#> [bytefactor]
#> [1] a b c a d
#> Levels: a b c d
small[1:3]
#> [bytefactor]
#> [1] a b c
#> Levels: a b c d

bitfactor

bitfactor() will choose an appropriate number of bits to store the given number of levels.

In the following example, there are 4 levels, so bitfactor() chooses to store each value in the factor in 2 bits.

tiny <- bitfactor(c('a', 'b', 'c', 'a', 'd'))
tiny
#> [bitfactor] 5 values @ ~2 bits/value = 1 integer(s)
#> [1] a b c a d
#> Levels: a b c d
tiny[1:3]
#> [bitfactor] 3 values @ ~2 bits/value = 1 integer(s)
#> [1] a b c
#> Levels: a b c d

In this next example, there are 100 levels in the factor, so 7 bits are needed to fully store all the levels

tiny <- bitfactor(sample(100, 10), levels=1:100)
tiny
#> [bitfactor] 10 values @ ~7 bits/value = 3 integer(s)
#>  [1] 24 77 87 57 33 21 18 2  11 62
#> 100 Levels: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 ... 100
tiny[4:6]
#> [bitfactor] 3 values @ ~7 bits/value = 1 integer(s)
#> [1] 57 33 21
#> 100 Levels: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 ... 100

Storing some DNA in bytefactor and bitfactor objects.

character factor bytefactor bitfactor
bits/value 64 32 8 2
total size 8 MB 4 MB 1 MB 270 kB
size reduction 2x 8x 30x
library(smallfactor)

#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Generate some random DNA
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
character_vector_dna <- sample(c('A', 'T', 'G', 'C'), 1e6, replace = TRUE)

#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Create a `factor` and a `smallfactor` using the same basic syntax
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
integer_factor <- factor    (character_vector_dna, levels = c('A', 'T', 'G', 'C'))
byte_factor    <- bytefactor(character_vector_dna, levels = c('A', 'T', 'G', 'C'))
bit_factor     <- bitfactor (character_vector_dna, levels = c('A', 'T', 'G', 'C'))

#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# `smallfactor` is approx 1/4 the size of the regular factor
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
lobstr::obj_size(character_vector_dna)
#> 8,000,272 B
lobstr::obj_size(integer_factor)
#> 4,000,688 B
lobstr::obj_size(byte_factor)
#> 1,000,696 B
lobstr::obj_size(bit_factor)
#> 267,704 B

Similar projects

Acknowledgements

  • R Core for developing and maintaining the language.
  • CRAN maintainers, for patiently shepherding packages onto CRAN and maintaining the repository