Comparison with {hash}

This vignette provides a comparison of {r2r} with the same-purpose CRAN package {hash}, which also offers an implementation of hash tables based on R environments. We first describe the features offered by both packages, and then perform some benchmark timing comparisons. The package versions referred to in this vignette are:

library(hash)
library(r2r)
packageVersion("hash")
#> [1] '2.2.6.4'
packageVersion("r2r")
#> [1] '0.1.2'

Features

Both {r2r} and {hash} hash tables are built on top of the R built-in environment data structure, and have thus a similar API. In particular, hash table objects have reference semantics for both packages. {r2r} hashtables are S3 class objects, whereas in {hash} the data structure is implemented as an S4 class.

Hash tables provided by r2r support arbitrary type keys and values, arbitrary key comparison and hash functions, and have customizable behaviour (either throw an exception or return a default value) upon query of a missing key.

In contrast, hash tables in hash currently support only string keys, with basic identity comparison (the hashing is performed automatically by the underlying environment objects); values can be arbitrary R objects. Querying missing keys through non-vectorized [[-subsetting returns the default value NULL, whereas queries through vectorized [-subsetting result in an error. On the other hand, hash also offers support for inverting hash tables (an experimental feature at the time of writing).

The table below summarizes the features of the two packages

Features supported by {r2r} and {hash}
Feature r2r hash
Basic data structure R environment R environment
Arbitrary type keys X
Arbitrary type values X X
Arbitrary hash function X
Arbitrary key comparison function X
Throw or return default on missing keys X
Hash table inversion X

Performance tests

We will perform our benchmark tests using the CRAN package microbenchmark.

library(microbenchmark)

Key insertion

We start by timing the insertion of:

N <- 1e4

random key-value pairs (with possible repetitions). In order to perform a meaningful comparison between the two packages, we restrict to string (i.e. length one character) keys. We can generate random keys as follows:

chars <- c(letters, LETTERS, 0:9)
random_keys <- function(n) paste0(
    sample(chars, n, replace = TRUE),
    sample(chars, n, replace = TRUE),
    sample(chars, n, replace = TRUE),
    sample(chars, n, replace = TRUE),
    sample(chars, n, replace = TRUE)
    )

set.seed(840)
keys <- random_keys(N)
values <- rnorm(N)

We test both the non-vectorized ([[<-) and vectorized ([<-) operators:

microbenchmark(
    `r2r_[[<-` = {
        for (i in seq_along(keys))
            m_r2r[[ keys[[i]] ]] <- values[[i]]
    },
    `r2r_[<-` = { m_r2r[keys] <- values },
    `hash_[[<-` = { 
        for (i in seq_along(keys))
            m_hash[[ keys[[i]] ]] <- values[[i]]
    },
    `hash_[<-` = m_hash[keys] <- values,
    
    times = 30, 
    setup = { m_r2r <- hashmap(); m_hash <- hash() }
)
#> Unit: milliseconds
#>       expr      min        lq      mean    median        uq      max neval
#>   r2r_[[<- 98.30017 129.21737 168.36815 168.26847 203.24199 258.1399    30
#>    r2r_[<- 69.21572  92.36945 139.38775 139.51036 170.67228 322.3156    30
#>  hash_[[<- 73.51051 111.01555 134.30033 127.65051 144.80263 239.5006    30
#>   hash_[<- 41.20582  69.54075  81.07606  84.44059  92.25313 140.8182    30

As it is seen, r2r and hash have comparable performances at the insertion of key-value pairs, with both vectorized and non-vectorized insertions, hash being somewhat more efficient in both cases.

Key query

We now test key query, again both in non-vectorized and vectorized form:

microbenchmark(
    `r2r_[[` = { for (key in keys) m_r2r[[ key ]] },
    `r2r_[` = { m_r2r[ keys ] },
    `hash_[[` = { for (key in keys) m_hash[[ key ]] },
    `hash_[` = { m_hash[ keys ] },
    
    times = 30,
    setup = { 
        m_r2r <- hashmap(); m_r2r[keys] <- values
        m_hash <- hash(); m_hash[keys] <- values
    }
)
#> Unit: milliseconds
#>     expr      min        lq      mean    median        uq       max neval
#>   r2r_[[ 96.07889 135.50293 185.64756 188.43839 233.46516 289.67895    30
#>    r2r_[ 92.25452 136.19839 171.05932 177.39769 213.41532 249.42809    30
#>  hash_[[ 11.32204  12.47591  17.06267  14.85390  21.46659  29.96053    30
#>   hash_[ 63.57955  76.25621 104.67501  98.25973 135.63418 158.75248    30

For non-vectorized queries, hash is significantly faster (by one order of magnitude) than r2r. This is likely due to the fact that the [[ method dispatch is handled natively by R in hash (i.e. the default [[ method for environments is used ), whereas r2r suffers the overhead of S3 method dispatch. This is confirmed by the result for vectorized queries, which is comparable for the two packages; notice that here a single (rather than N) S3 method dispatch occurs in the r2r timed expression.

As an additional test, we perform the benchmarks for non-vectorized expressions with a new set of keys:

set.seed(841)
new_keys <- random_keys(N)
microbenchmark(
    `r2r_[[_bis` = { for (key in new_keys) m_r2r[[ key ]] },
    `hash_[[_bis` = { for (key in new_keys) m_hash[[ key ]] },
    
    times = 30,
    setup = { 
        m_r2r <- hashmap(); m_r2r[keys] <- values
        m_hash <- hash(); m_hash[keys] <- values
    }
)
#> Unit: milliseconds
#>         expr      min       lq      mean    median        uq       max neval
#>   r2r_[[_bis 71.58680 96.61528 127.11334 125.18600 159.67464 226.21784    30
#>  hash_[[_bis 11.47926 12.57271  18.66148  18.26924  24.03328  32.31627    30

The results are similar to the ones already commented. Finally, we test the performances of the two packages in checking the existence of keys (notice that here has_key refers to r2r::has_key, whereas has.key is hash::has.key):

set.seed(842)
mixed_keys <- sample(c(keys, new_keys), N)
microbenchmark(
    r2r_has_key = { for (key in mixed_keys) has_key(m_r2r, key) },
    hash_has_key = { for (key in new_keys) has.key(key, m_hash) },
    
    times = 30,
    setup = { 
        m_r2r <- hashmap(); m_r2r[keys] <- values
        m_hash <- hash(); m_hash[keys] <- values
    }
)
#> Unit: milliseconds
#>          expr       min        lq     mean   median       uq      max neval
#>   r2r_has_key  74.30419  85.32149 106.2206 108.9638 123.8719 141.9533    30
#>  hash_has_key 192.51283 234.37185 304.9381 300.7907 374.1574 439.1151    30

The results are comparable for the two packages, r2r being slightly more performant in this particular case.

Key deletion

Finally, we test key deletion. In order to handle name collisions, we will use delete() (which refers to r2r::delete()) and del() (which refers to hash::del()).

microbenchmark(
    r2r_delete = { for (key in keys) delete(m_r2r, key) },
    hash_delete = { for (key in keys) del(key, m_hash) },
    hash_vectorized_delete = { del(keys, m_hash) },
    
    times = 30,
    setup = { 
        m_r2r <- hashmap(); m_r2r[keys] <- values
        m_hash <- hash(); m_hash[keys] <- values
    }
)
#> Unit: milliseconds
#>                    expr        min         lq       mean     median        uq
#>              r2r_delete 128.781056 149.410034 201.068762 197.595870 239.84616
#>             hash_delete  65.164356  82.383820 112.735360 114.210446 136.74302
#>  hash_vectorized_delete   2.870008   3.208157   3.860566   3.677855   4.40237
#>         max neval
#>  291.472659    30
#>  196.634761    30
#>    6.294353    30

The vectorized version of hash significantly outperforms the non-vectorized versions (by roughly two orders of magnitude in speed). Currently, r2r does not support vectorized key deletion 1.

Conclusions

The two R packages r2r and hash offer hash table implementations with different advantages and drawbacks. r2r focuses on flexibility, and has a richer set of features. hash is more minimal, but offers superior performance in some important tasks. Finally, as a positive note for both parties, the two packages share a similar API, making it relatively easy to switch between the two, according to the particular use case needs.


  1. This is due to complications introduced by the internal hash collision handling system of r2r.↩︎