Performance regression with .N and :=

[svraka]

On the latest Github master dt[, N = .N, by = id] became much slower than on the latest CRAN version. The output below is from an M1 Mac but the issue can be reproduced on x64 Linux and Windows as well. I included tests for all optimization levels. I haven't found similar issues with other optimized functions.

CRAN (1.14.2)

library(data.table)
options(datatable.verbose = TRUE)

N <- 1e8
n <- 1e6

set.seed(1L)
dt <- data.table(g = sample(seq_len(n), N, TRUE),
                 x = runif(N),
                 key = "g")
#> forder.c received 100000000 rows and 2 columns
#> forder took 2.256 sec
#> reorder took 1.934s elapsed (4.338s cpu)

options(datatable.optimize = 0L)
dt_mod <- copy(dt)
dt_mod[, N := .N, by = g]
#> Detected that j uses these columns: <none> 
#> Finding groups using uniqlist on key ... 0.115s elapsed (0.097s cpu) 
#> Finding group sizes from the positions (can be avoided to save RAM) ... 0.004s elapsed (0.003s cpu) 
#> All optimizations are turned off
#> Making each group and running j (GForce FALSE) ... 
#>   memcpy contiguous groups took 0.310s for 1000000 groups
#>   eval(j) took 0.300s for 1000000 calls
#> 1.412s elapsed (0.630s cpu)

options(datatable.optimize = 1L)
dt_mod <- copy(dt)
dt_mod[, N := .N, by = g]
#> Detected that j uses these columns: <none> 
#> Finding groups using uniqlist on key ... 0.112s elapsed (0.099s cpu) 
#> Finding group sizes from the positions (can be avoided to save RAM) ... 0.003s elapsed (0.003s cpu) 
#> lapply optimization is on, j unchanged as '.N'
#> Making each group and running j (GForce FALSE) ... 
#>   memcpy contiguous groups took 0.312s for 1000000 groups
#>   eval(j) took 0.300s for 1000000 calls
#> 1.767s elapsed (0.610s cpu)

options(datatable.optimize = 2L)
dt_mod <- copy(dt)
dt_mod[, N := .N, by = g]
#> Detected that j uses these columns: <none> 
#> Finding groups using uniqlist on key ... 0.104s elapsed (0.096s cpu) 
#> Finding group sizes from the positions (can be avoided to save RAM) ... 0.004s elapsed (0.003s cpu) 
#> lapply optimization is on, j unchanged as '.N'
#> Making each group and running j (GForce FALSE) ... 
#>   memcpy contiguous groups took 0.311s for 1000000 groups
#>   eval(j) took 0.298s for 1000000 calls
#> 1.442s elapsed (0.613s cpu)

options(datatable.optimize = 3L)
dt_mod <- copy(dt)
dt_mod[, N := .N, by = g]
#> Detected that j uses these columns: <none> 
#> Finding groups using uniqlist on key ... 0.114s elapsed (0.098s cpu) 
#> Finding group sizes from the positions (can be avoided to save RAM) ... 0.003s elapsed (0.003s cpu) 
#> lapply optimization is on, j unchanged as '.N'
#> Making each group and running j (GForce FALSE) ... 
#>   memcpy contiguous groups took 0.308s for 1000000 groups
#>   eval(j) took 0.296s for 1000000 calls
#> 1.364s elapsed (0.595s cpu)

^{Created on 2022-07-25 by the reprex package (v2.0.1)}

Session info

sessionInfo()
#> R version 4.2.1 (2022-06-23)
#> Platform: aarch64-apple-darwin20 (64-bit)
#> Running under: macOS Monterey 12.5
#> 
#> Matrix products: default
#> BLAS:   /Library/Frameworks/R.framework/Versions/4.2-arm64/Resources/lib/libRblas.0.dylib
#> LAPACK: /Library/Frameworks/R.framework/Versions/4.2-arm64/Resources/lib/libRlapack.dylib
#> 
#> locale:
#> [1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8
#> 
#> attached base packages:
#> [1] stats     graphics  grDevices utils     datasets  methods   base     
#> 
#> other attached packages:
#> [1] data.table_1.14.2
#> 
#> loaded via a namespace (and not attached):
#>  [1] digest_0.6.29   withr_2.5.0     magrittr_2.0.3  reprex_2.0.1   
#>  [5] evaluate_0.15   highr_0.9       stringi_1.7.8   rlang_1.0.4    
#>  [9] cli_3.3.0       fs_1.5.2        rmarkdown_2.14  tools_4.2.1    
#> [13] stringr_1.4.0   glue_1.6.2      xfun_0.31       yaml_2.3.5     
#> [17] fastmap_1.1.0   compiler_4.2.1  htmltools_0.5.3 knitr_1.39

Github (master)

library(data.table)
options(datatable.verbose = TRUE)

N <- 1e8
n <- 1e6

set.seed(1L)
dt <- data.table(g = sample(seq_len(n), N, TRUE),
                 x = runif(N),
                 key = "g")
#> forder.c received 100000000 rows and 2 columns
#> forder took 2.672 sec
#> reorder took 1.965s elapsed (4.403s cpu)

options(datatable.optimize = 0L)
dt_mod <- copy(dt)
dt_mod[, N := .N, by = g]
#> Argument 'by' after substitute: g
#> Detected that j uses these columns: <none>
#> Finding groups using uniqlist on key ... 0.174s elapsed (0.105s cpu) 
#> Finding group sizes from the positions (can be avoided to save RAM) ... 0.014s elapsed (0.003s cpu) 
#> All optimizations are turned off
#> Making each group and running j (GForce FALSE) ... 
#>   memcpy contiguous groups took 0.260s for 1000000 groups
#>   eval(j) took 0.249s for 1000000 calls
#> 2.103s elapsed (1.489s cpu)

options(datatable.optimize = 1L)
dt_mod <- copy(dt)
dt_mod[, N := .N, by = g]
#> Argument 'by' after substitute: g
#> Detected that j uses these columns: <none>
#> Finding groups using uniqlist on key ... 0.116s elapsed (0.099s cpu) 
#> Finding group sizes from the positions (can be avoided to save RAM) ... 0.004s elapsed (0.003s cpu) 
#> lapply optimization is on, j unchanged as '.N'
#> Making each group and running j (GForce FALSE) ... 
#>   memcpy contiguous groups took 0.263s for 1000000 groups
#>   eval(j) took 0.249s for 1000000 calls
#> 2.294s elapsed (1.479s cpu)

options(datatable.optimize = 2L)
dt_mod <- copy(dt)
dt_mod[, N := .N, by = g]
#> Argument 'by' after substitute: g
#> Detected that j uses these columns: <none>
#> Finding groups using uniqlist on key ... 0.096s elapsed (0.094s cpu) 
#> Finding group sizes from the positions (can be avoided to save RAM) ... 0.004s elapsed (0.003s cpu) 
#> lapply optimization is on, j unchanged as '.N'
#> Making each group and running j (GForce FALSE) ... 
#>   memcpy contiguous groups took 0.262s for 1000000 groups
#>   eval(j) took 0.249s for 1000000 calls
#> 2.096s elapsed (1.468s cpu)

options(datatable.optimize = 3L)
dt_mod <- copy(dt)
dt_mod[, N := .N, by = g]
#> Argument 'by' after substitute: g
#> Detected that j uses these columns: <none>
#> Finding groups using uniqlist on key ... 0.114s elapsed (0.098s cpu) 
#> Finding group sizes from the positions (can be avoided to save RAM) ... 0.003s elapsed (0.004s cpu) 
#> lapply optimization is on, j unchanged as '.N'
#> Making each group and running j (GForce FALSE) ... 
#>   memcpy contiguous groups took 0.261s for 1000000 groups
#>   eval(j) took 0.247s for 1000000 calls
#> 2.048s elapsed (1.464s cpu)

^{Created on 2022-07-26 by the reprex package (v2.0.1)}

Session info

sessionInfo()
#> R version 4.2.1 (2022-06-23)
#> Platform: aarch64-apple-darwin20 (64-bit)
#> Running under: macOS Monterey 12.5
#> 
#> Matrix products: default
#> BLAS:   /Library/Frameworks/R.framework/Versions/4.2-arm64/Resources/lib/libRblas.0.dylib
#> LAPACK: /Library/Frameworks/R.framework/Versions/4.2-arm64/Resources/lib/libRlapack.dylib
#> 
#> locale:
#> [1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8
#> 
#> attached base packages:
#> [1] stats     graphics  grDevices utils     datasets  methods   base     
#> 
#> other attached packages:
#> [1] data.table_1.14.3
#> 
#> loaded via a namespace (and not attached):
#>  [1] digest_0.6.29   withr_2.5.0     magrittr_2.0.3  reprex_2.0.1   
#>  [5] evaluate_0.15   highr_0.9       stringi_1.7.8   rlang_1.0.4    
#>  [9] cli_3.3.0       fs_1.5.2        rmarkdown_2.14  tools_4.2.1    
#> [13] stringr_1.4.0   glue_1.6.2      xfun_0.31       yaml_2.3.5     
#> [17] fastmap_1.1.0   compiler_4.2.1  htmltools_0.5.3 knitr_1.39

[tdhock]

I did a git bisect and I found that the slowdown started after merging #4491 which was done by @jangorecki so maybe you have some idea how to fix?
In detail here are the commits near where the slowdown occured

and here are the timings,

The plot above shows that #4491 is just as slow as master, whereas the two commits immediate before that are significantly faster, which implies that PR is responsible for the slowdown.

[ben-schwen]

Well if it was introduced by #4491 then it might be the same issue with snprintf() as in #5371

[jangorecki]

I think it is exactly what Ben noticed. Fix is simple because this part of code is responsible for verbose output, not the actual computation.

[tdhock]

Is there a commit with that supposed simple fix that I can test?
To generate that timings figure I did not use verbose output, which suggests the problem is something else.

[jangorecki]

Those messages are generated always and just displayed with verbose output. It may have been introduced in e793f53
In the other issue I asked Matt about preferred solution from few options proposed there.

[tdhock]

I did another git bisect, using only the commits in #4491.

The figure above confirms that the slowdown started happening in that commit, e793f53 which was attributed to @mattdowle
BTW these figures are made by timing the following code for different versions and different values of N,

  N = 10^6,
    n <- N/100
    set.seed(1L)
    dt <- data.table(
      g = sample(seq_len(n), N, TRUE),
      x = runif(N),
      key = "g")
    dt_mod <- copy(dt)
    data.table:::`[.data.table`(dt_mod, , N := .N, by = g)