One feature of canaper is the ability to use parallel
computing (running calculations on multiple CPUs simultaneously) to
speed up analysis. The parallel computing is used during the
randomizations carried out by cpr_rand_test(), since this
function involves calculating the same values on many random replicates.
This vignette shows how and when to use parallel computing to speed up
cpr_rand_test().
(This vignette assumes a basic understanding of CANAPE, community data matrices, and randomizations. If you aren’t familiar with any of these, you should probably see the the CANAPE example vignette first).
Let’s get started by loading the packages used in this vignette.
library(canaper) # This package
library(tictoc) # For timing
library(future) # For parallel computing
# Set seed for random number generator for reproducible results
set.seed(12345)canaper uses the future package to handle
parallel computing. In future, specification of sequential
(i.e., no parallel computing) vs. parallel computing, and the number of
CPUs (i.e., “cores”) to use in parallel is specified
outside of other functions. This is easiest to see with
an example.
First, let’s run an analysis in default sequential mode (without
parallel computing). I’ll use the tictoc package to time
how long it takes to run.
tic()
biod_res_seq <- cpr_rand_test(
biod_example$comm, biod_example$phy,
null_model = "swap", n_reps = 50
)
#> Warning: Abundance data detected. Results will be the same as if using presence/absence data (no abundance weighting is used).
toc()
#> 5.131 sec elapsedSince we have specified 50 random replicates and are not using
parallel computing, cpr_rand_test() calculated the various
phylogenetic diversity metrics for each of the 50 replicates one at a
time.
Before trying the parallel version, let’s check how many CPUs are available to use:
availableCores()
#> system
#> 6OK, we have verified that there are multiple cores available for parallel computing.
To enable parallel computing, just add one line before
cpr_rand_test():
plan(multisession, workers = 2) 1. Here, the
multisession, workers = 2 part is telling
future that we want to use 2 CPUs in parallel on our local
machine. See future::plan() for other options. Otherwise,
everything is the same.
# Set up parallel computing with 2 CPUs
plan(multisession, workers = 2)
tic()
biod_res_par <- cpr_rand_test(
biod_example$comm, biod_example$phy,
null_model = "swap", n_reps = 50
)
#> Warning: Abundance data detected. Results will be the same as if using presence/absence data (no abundance weighting is used).
toc()
#> 9.136 sec elapsed
# Change back to default sequential mode
plan(sequential)This time, the calculations were carried out in 2 batches in parallel.
But there is no significant improvement in processing time2. What is going on here?
Although it may seem to always be a good idea to speed things up by using parallel computing, this is not the case. There is some computational overhead involved in splitting the job across multiple processes, coordinating those processes, and putting everything back together again.
If your dataset is small, this overhead may outweigh simply running
the analysis sequentially. That is the case with the
biod_example data. Let’s check the size of this
dataset:
# dim() returns number of rows, then columns
dim(biod_example$comm)
#> [1] 127 31The biod_example dataset is small because it is entirely
made-up and used only for testing code (and we want tests to run
quickly).
Let’s see how that compares with another dataset included in
canaper, the acacia dataset. The
acacia dataset is “real-life” data of the genus
Acacia in Australia:
# dim() returns number of rows, then columns
dim(acacia$comm)
#> [1] 3037 508Quite a bit larger!
Let’s see how parallel computing works on the acacia
dataset:
plan(sequential)
tic()
acacia_res_seq <- cpr_rand_test(
acacia$comm, acacia$phy,
null_model = "curveball", n_reps = 100
)
#> Warning: Abundance data detected. Results will be the same as if using presence/absence data (no abundance weighting is used).
#> Warning: 'comm' is > 95% absences (zeros). Be sure that 'n_reps' and 'n_iterations' are sufficiently large to ensure adequate mixing of random communities
#> Warning: Dropping tips from the tree because they are not present in the community data:
#> Pararchidendron_pruinosum, Paraserianthes_lophantha
toc()
#> 107.131 sec elapsed# Run cpr_rand_test() in parallel with 2 CPUs
plan(multisession, workers = 2)
tic()
acacia_par_seq <- cpr_rand_test(
acacia$comm, acacia$phy,
null_model = "curveball", n_reps = 100
)
#> Warning: Abundance data detected. Results will be the same as if using presence/absence data (no abundance weighting is used).
#> Warning: 'comm' is > 95% absences (zeros). Be sure that 'n_reps' and 'n_iterations' are sufficiently large to ensure adequate mixing of random communities
#> Warning: Dropping tips from the tree because they are not present in the community data:
#> Pararchidendron_pruinosum, Paraserianthes_lophantha
toc()
#> 90.475 sec elapsed
plan(sequential)And now we start to see the performance improvements that be can be gained from parallel computing!3
If you’d like to track the progress of cpr_rand_test()
in real time, you can enable a progress bar. There are two ways to do so
(neither of these will show up on the webpage, but do try it at home!).
Similar to parallelization, this is done outside of the actual
function.
One way is to add progressr::handlers(global = TRUE)
before cpr_rand_test():
progressr::handlers(global = TRUE)
biod_res_long <- cpr_rand_test(
biod_example$comm, biod_example$phy,
null_model = "swap", n_reps = 500
)The other way is to place cpr_rand_test() inside the
progressr::with_progress() function:
progressr::with_progress(
biod_res_long <- cpr_rand_test(
biod_example$comm, biod_example$phy,
null_model = "swap", n_reps = 500
)
)This vignette shows how easy it is to enable parallel computing in
canaper, and when it makes sense to do so. I hope it helps
your analyses run faster!
Of course, the number of workers should be no greater than the number of available CPUs.↩︎
Results vary each time this vignette is generated, but it is usually about the same; sometimes, the parallel version is even slower.↩︎
The settings used here are chosen to demonstrate performance gains from parellelization with short overall computation time, not for accurately calculating CANAPE.↩︎