I am currently using furrr to create a more organized execution of my model. I use a data.frame to pass parameters to a function in a orderly way, and then using the furrr::future_map() to map a function across all the parameters. The function works flawlessly when using the sequential and multicore futures on my local Machine (OSX).
Now, I want to test my code creating my own cluster of AWS instances (just as shown here).
I created a function using the linked article code:
make_cluster_ec2 <- function(public_ip){
ssh_private_key_file <- Sys.getenv('PEM_PATH')
github_pac <- Sys.getenv('PAC')
cl_multi <- future::makeClusterPSOCK(
workers = public_ip,
user = "ubuntu",
rshopts = c(
"-o", "StrictHostKeyChecking=no",
"-o", "IdentitiesOnly=yes",
"-i", ssh_private_key_file
),
rscript_args = c(
"-e", shQuote("local({p <- Sys.getenv('R_LIBS_USER'); dir.create(p, recursive = TRUE, showWarnings = FALSE); .libPaths(p)})"),
"-e", shQuote("install.packages('devtools')"),
"-e", shQuote(glue::glue("devtools::install_github('user/repo', auth_token = '{github_pac}')"))
),
dryrun = FALSE)
return(cl_multi)
}
Then, I create the cluster object and then check that is connected to the right instance
public_ids <- c('public_ip_1', 'public_ip_2')
cls <- make_cluster_ec2(public_ids)
f <- future(Sys.info())
And when I print f I get the specs of one of my remote instances, which indicates the socket is correctly connected:
> value(f)
sysname
"Linux"
release
"4.15.0-1037-aws"
version
"#39-Ubuntu SMP Tue Apr 16 08:09:09 UTC 2019"
nodename
"ip-xxx-xx-xx-xxx"
machine
"x86_64"
login
"ubuntu"
user
"ubuntu"
effective_user
"ubuntu"
But when I run my code using my cluster plan:
plan(list(tweak(cluster, workers = cls), multisession))
parameter_df %>%
mutate(model_traj = furrr::future_pmap(list('lat' = latitude,
'lon' = longitude,
'height' = stack_height,
'name_source' = facility_name,
'id_source' = facility_id,
'duration' = duration,
'days' = seq_dates,
'daily_hours' = daily_hours,
'direction' = 'forward',
'met_type' = 'reanalysis',
'met_dir' = here::here('met'),
'exec_dir' = here::here("Hysplit4/exec"),
'cred'= list(creds)),
dirtywind::hysplit_trajectory,
.progress = TRUE)
)
I get the following error:
Error in file(temp_file, "a") : cannot open the connection
In addition: Warning message:
In file(temp_file, "a") :
cannot open file '/var/folders/rc/rbmg32js2qlf4d7cd4ts6x6h0000gn/T//RtmpPvdbV3/filecf23390c093.txt': No such file or directory
I can not figure out what is happening under the hood, and I can not traceback() the error either from my remote machines. I have test the connection with the examples in the article and things seem to run correctly. I am wondering why is trying to create a tempdir during the execution. What am I missing here?
(This is also an issue in the furrr repo)
Disable the progress bar, i.e. don't specify .progress = TRUE.
This is because .progress = TRUE assumes your R workers can write to a temporary file that the main R process created. This is typically only possible when you parallelize on the same machine.
A smaller example of this error is:
library(future)
## Set up a cluster with one worker running on another machine
cl <- makeClusterPSOCK(workers = "node2")
plan(cluster, workers = cl)
y <- furrr::future_map(1:2, identity, .progress = FALSE)
str(y)
## List of 2
## $ : int 1
## $ : int 2
y <- furrr::future_map(1:2, identity, .progress = TRUE)
## Error in file(temp_file, "a") : cannot open the connection
## In addition: Warning message:
## In file(temp_file, "a") :
## cannot open file '/tmp/henrik/Rtmp1HkyJ8/file4c4b864a028ac.txt': No such file or directory
If you want to get near-live progress updates, you can instead use the progressr package. It works with the Futureverse, including furrr. See https://progressr.futureverse.org/articles/progressr-intro.html#future_map---parallel-purrrmap for an example.
EDIT 2023-07-06: Mention progressr.