slurm_apply a RefClass method from within a RefClass method - r

Question

EDIT: New version of rslurm makes the solution very easy. See my answer below.
Apologies for the somewhat longer than desired MWE, and a title that I realize after submitting the question may be needlessly complicated. I believe the real issue is getting the environment of a RefClass object into rslurm::slurm_apply.
MWE
Here I define a toy reference class called BankAccount. It has two fields and two methods.
The fields are transactions, a list of all transactions associated with the account and suspicion_threshold the value above which the bank will investigate the transaction.
The two methods are is_suspicious which compares the transactions with the suspicion_threshold on the local machine and is_suspicious_slurm, which uses rslurm::slurm_apply to spread many calls to is_suspicious over a cluster of computers managed by SLURM. You can imagine if there were many transactions or if the is_suspicious function were more complex, this might be necessary.
So, here's the setup
BankAccount <- setRefClass(
Class = 'BankAccount',
fields = list(
transactions = 'numeric',
suspicion_threshold = 'numeric'
)
)
BankAccount$methods(
is_suspicious = function(start_idx = 1, stop_idx = length(transactions)) {
return(start_idx + which(transactions[start_idx:stop_idx] > suspicion_threshold) - 1)
}
)
BankAccount$methods(
is_suspicious_slurm = function(num_nodes) {
usingMethods(is_suspicious)
t <- length(transactions)
t_per_n <- floor(t/num_nodes)
starts <- seq(from = 1, length.out = num_nodes, by = t_per_n)
stops <- seq(from = t_per_n, length.out = num_nodes, by = t_per_n)
stops[num_nodes] <- t
sjob <- rslurm::slurm_apply(f = is_suspicious,
params = data.frame(start_idx = starts,
stop_idx = stops),
nodes = num_nodes,
add_objects = .self)
results_list <- rslurm::get_slurm_out(slr_job = sjob,
outtype = "raw",
wait = TRUE)
return(unlist(results_list))
}
)
Now, on my local machine I can run:
library(RCexampleforSE)
set.seed(27599)
b <- BankAccount$new()
b$transactions <- rnorm(n = 500)
b$suspicion_threshold <- 2
b$is_suspicious()
b$is_suspicious_slurm(num_nodes = 3)
and it works as expected:
62 103 155 171 182 188 297 398 493 499
If I run:
b$is_suspicious_slurm(num_nodes = 3)
I get an error, since my personal computer is not connected to a SLURM cluster.
sh: squeue: command not found
Cannot submit; no SLURM workload manager on path
Submission scripts output in directory _rslurm_13ba46e3c70b0
Error in rslurm::get_slurm_out(slr_job = sjob, outtype = "raw", wait = TRUE):
slr_job has not been submitted
If I logon to my university cluster, which uses SLURM, and run the same script, the setup and local methods work just as they did on my personal computer. When I run:
b$is_suspicious_slurm(num_nodes = 3)
it sends jobs to the cluster, as hoped for:
Submitted batch job 6363868
But these jobs error immediately with the following error message in slurm_0.out, slurm_1.out, and slurm_2.out:
Error in attr(, "mayCall") : argument 1 is empty
Execution halted
Thoughts and Attempts
I figure the job probably needs, but doesn't have available, the BankAccount object. So I tried passing it in as add_objects parameter to rslurm::slurm_apply:
sjob <- rslurm::slurm_apply(f = is_suspicious,
params = data.frame(start_idx = starts,
stop_idx = stops),
nodes = num_nodes,
add_objects = .self)
I also tried it in quotes and inside eval(), neither of which worked.
How can I make the object accessible to the worker jobs created with rslurm::slurm_apply?

Answer 1

Version 0.4.0 of rslurm completely solved this problem.
Define is_suspicious_slurm() as:
BankAccount$methods(
is_suspicious_slurm = function(num_nodes) {
usingMethods(is_suspicious)
t <- length(transactions)
t_per_n <- floor(t/num_nodes)
starts <- seq(from = 1, length.out = num_nodes, by = t_per_n)
stops <- seq(from = t_per_n, length.out = num_nodes, by = t_per_n)
stops[num_nodes] <- t
sjob <- rslurm::slurm_apply(f = is_suspicious,
params = data.frame(start_idx = starts,
stop_idx = stops),
nodes = num_nodes)
results_list <- rslurm::get_slurm_out(slr_job = sjob,
outtype = "raw",
wait = TRUE)
return(unlist(results_list))
}
)
The only change is that in the call to rslurm::slurm_apply, the add_objects parameter is not specified. It does not need to be specified because as #Ian pointed out:
"...you don't need to pass self at all when slurm_apply sends the serialized function, which appears to include both ".self" and "transactions" in the enclosing environment."

Answer 2

EDIT: OP's answer is all you need to know.
The add_objects parameter is used for passing a character vector, not the objects themselves. All the objects are then saved in one RData file, assuming they can be found by name. In theory, you should be able to use add_objects = c('.self') within your method definition.
The key here is, "assuming they can be found". I will edit this post once a pending update to the rslurm package (which should make that finding more successful) is released.
Be very careful passing objects to cluster nodes: they do not come back. Not only will any side effects be lost, there's no inter-node communication implemented by rslurm.
Also be careful with which :) Your is_suspicious method will be wrong for arguments that don't start at 1. Try this version:
BankAccount$methods(
is_suspicious = function(i = 1:length(transactions)) {
idx <- which(transactions[i] > suspicion_threshold)
i[idx]
}
)