[R-package] Quick question about num of thread

[issactoast]

Hi, Thank you for making lightGBM in R!

I am using LightGBM in R and have a quick question about the num_thread.

According to the manual, the number of threads is the physical core of CPU. But usually what I have seen in my R code, set the num thread is equal to num_thread - 1 such as

cores <- parallel::detectCores() -1
cores

So if we have 4 core, use 3 for parallel and 1 for the controller. Is this applied to LightGBM too? So if I have 4 physical core of CPU and set the number of thread as 3?

[jameslamb]

Thanks for using LightGBM!

Assuming that there are not other processes making heavy use of the available CPUs, you will get the best performance by setting num_thread equal to the total number of real CPU cores. There is not really a concept of a "controller" in parallel training with LightGBM.

You could use something like this to test the relative speedup from different settings of num_threads:

library(lightgbm)
library(microbenchmark)
library(nycflights13)

data(flights, package = "nycflights13")
flights <- as.data.frame(flights)

dtrain <- lgb.Dataset(
    as.matrix(
        flights[, c("year", "sched_dep_time", "distance", "hour", "minute")]
    )
    , label = flights[, "dep_delay"]
    , free_raw_data = FALSE
    , max_bin = 350
)

num_cores <- parallel::detectCores()

for (num_thread in c(num_cores - 1, num_cores)) {
    print(paste0("num_thread: ", num_thread))
    print(
        microbenchmark::microbenchmark({
            lgb.train(
                params = list(
                    num_thread = num_thread
                    , objective = "regression_l2"
                    , num_leaves = 31L
                    , max_depth = 8L
                    , learning_rate = 0.01
                    , min_data_in_leaf = 1
                )
                , data = dtrain
                , nrounds = 1000L
                , verbose = -1L
            )
        }, times = 5, unit = "s")
    )
}

I installed {lightgbm} 3.2.1 on my Mac tonight with install.packages("lightgbm", type = "source"), and with that version I got the following results from the code above. I have two 2-core CPUs on this machine.

# num_threads = 3

     min       lq     mean   median       uq    max neval
 5.23764 5.273885 5.497161 5.331118 5.625063 6.0181     5

# num_threads = 4
      min       lq     mean   median       uq      max neval
 4.582549 4.762857 4.907298 4.837003 4.925407 5.428671     5

Your specific results will vary based on your specific dataset and the other learning parameter values you set.

[issactoast]

Thank you for the clarification! Just want to make it clear, we need to use num_cores <- parallel::detectCores(logical=FALSE) so that the num_cores is equal to the physical core. I used the same code that you have and it confirms that setting the num_thread = num of physical core is faster. Thank you!

library(lightgbm)
#> Loading required package: R6
library(microbenchmark)
library(nycflights13)

data(flights, package = "nycflights13")
flights <- as.data.frame(flights)

dtrain <- lgb.Dataset(
    as.matrix(
        flights[, c("year", "sched_dep_time", "distance", "hour", "minute")]
    )
    , label = flights[, "dep_delay"]
    , free_raw_data = FALSE
    , max_bin = 350
)

num_cores <- parallel::detectCores(logical = FALSE)

for (num_thread in c(num_cores, num_cores * 2)) {
    print(paste0("num_thread: ", num_thread))
    print(
        microbenchmark::microbenchmark({
            lgb.train(
                params = list(
                    num_thread = num_thread
                    , objective = "regression_l2"
                    , num_leaves = 31L
                    , max_depth = 8L
                    , learning_rate = 0.01
                    , min_data_in_leaf = 1
                )
                , data = dtrain
                , nrounds = 1000L
                , verbose = -1L
            )
        }, times = 5, unit = "s")
    )
}
#> [1] "num_thread: 10"
#> Unit: seconds
#>                                                                                                                                                                                                                                            expr
#>  {     lgb.train(params = list(num_thread = num_thread, objective = "regression_l2",          num_leaves = 31L, max_depth = 8L, learning_rate = 0.01,          min_data_in_leaf = 1), data = dtrain, nrounds = 1000L,          verbose = -1L) }
#>       min       lq     mean   median       uq      max neval
#>  2.522868 2.546463 2.593207 2.563418 2.646491 2.686795     5
#> [1] "num_thread: 20"
#> Unit: seconds
#>                                                                                                                                                                                                                                            expr
#>  {     lgb.train(params = list(num_thread = num_thread, objective = "regression_l2",          num_leaves = 31L, max_depth = 8L, learning_rate = 0.01,          min_data_in_leaf = 1), data = dtrain, nrounds = 1000L,          verbose = -1L) }
#>       min       lq    mean   median       uq      max neval
#>  4.536631 4.673258 4.67483 4.682869 4.687078 4.794311     5

^{Created on 2021-04-18 by the reprex package (v1.0.0)}

[jameslamb]

Ah yes, you are absolutely right! Are you interested in contributing a change to the documentation? I think others would benefit from that note.

It would just be updating

LightGBM/R-package/R/lightgbm.R

Lines 95 to 97 in 7ea2bc4

	#' \item{\code{num_threads}: Number of threads for LightGBM. For the best speed, set this to
	#' the number of real CPU cores, not the number of threads (most
	#' CPU using hyper-threading to generate 2 threads per CPU core).}

to say something like

the number of real CPU cores (\code{parallel::detectCores(logical = FALSE)})

And then re-generating the documentation files with commands like this:

sh build-cran-package.sh
R CMD INSTALL --with-keep.source lightgbm_*.tar.gz
cd R-package
Rscript -e "roxygen2::roxygenize(load = 'installed')"

[issactoast]

@jameslamb Sure! I will do that, thanks!

[jameslamb]

Great, thanks so much! Let me know if you run into any issues.

[github-actions]

This issue has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.