num.threads not respected by ranger::ranger function

[alexzwanenburg]

I implemented the ranger::ranger function within a loop that is externally parallelised (using parallel::parLapply). To avoid competition across processes, I set num.threads to 1. However, each process is still attempting to maximise CPU use, which indicates that more than one thread is created.

When I disable external parallelisation, ranger still tries to maximise CPU use, despite num.threads=1.

My call to the ranger::ranger function is shown below:

model_obj      <- ranger::ranger(formula, data=dt,
                                     num.trees = 2^param$n_tree,
                                     sample.fraction = param$sample_size,
                                     mtry = ceiling(param$m_try*length(feature_cols)),
                                     min.node.size = param$node_size,
                                     alpha = param$alpha,
                                     splitrule = param$split_rule,
                                     importance = vimp_method,
                                     probability = fit_probability,
                                     num.threads = 1)

In the above code snippet, param is a list with model hyper-parameters, and vimp_method and fit_probability are likewise provided. These parameters should not affect threading.

I tried num.threads=1L, but this did not have any different effect.

The operating system is Windows 7 SP1 Build 7601, running on Intel i7-3770 CPU with 4 (8) cores and 32 GB of RAM. The version of Ranger is 0.10.1, with R 3.5.0.

If I am missing something obvious please let me know.

[alexzwanenburg]

Current usage is to predict a multinomial outcome. This outcome is a factor with four levels. fit_probability=TRUE for this case. The treetype of the model is Probability estimation.

[mnwright]

Are you sure it actually uses more than once core? In the current implementation there is still one child process started to grow the trees but the main thread should idle during that time.

We should change this behaviour and don't spawn a new thread for num.threads=1, but I wanted to wait for the old RTools to vanish to clean up that part, see also #41 (now over two years old...).

Nevertheless, it shouldn't use more than one core. Does it?

[alexzwanenburg]

I have tried to replicate the error using a simple example on the same data set. However, here I observe the expected behaviour, i.e. only one core is being used.

require(data.table)

# Read table. We are trying to predict the vote column
dt <- as.data.table(carData::Chile)

# Remove NA entries from the data table 
dt <- dt[!is.na(vote) & !is.na(age) & !is.na(education) & !is.na(income) & !is.na(statusquo)]

# This works as it should. Only one core is used.
model_obj <- ranger::ranger(vote ~ ., data=dt, num.trees=10000, num.threads = 1, probability=TRUE)

# This also works as it should. Only one core is used.
for(ii in 1:100) {
  model_obj <- ranger::ranger(vote ~ ., data=dt, num.trees=500, num.threads = 1, probability=TRUE)
}

After this example, I actually traced the problem to a mistake on my side. Directly after building the random forest, I use it to predict (ranger:::predict.ranger) outcomes for a new data set. This happens repeatedly. However, I did not set num.threads on the predict function, which caused it to multi-thread. The issue was resolved by setting num.threads=1 while performing the prediction.

Thanks for the help!

[alexzwanenburg]

This issue may be closed.