Zombie fix attempts

[paul-buerkner]

In this PR, I will try several things to fix #1658. @avehtari or @n-kall I would be gratefull if you could check it out on your systems because I currently don't have access to Ubuntu.

List of things tried out in this PR so far:

• copy the plapply code directly into log_lik (don't call plapply itself anymore)
• always use PSOCK clusters even on unix systems

[n-kall]

Edit: I don't get zombies on my ubuntu machine with the current master branch.

I think that first try fixed it. I don't seem to get zombies with the minimal example anymore. @avehtari how about you?

[avehtari]

Using the github master I didn't get zombies with the minimal example, but after running at least 333 lines of brms_demo.R I get zombies when calling brms::log_lik() also with this PR

[paul-buerkner]

thanks! is any of the code executed in parallel? and where can I found the brms_demo.R? Aki Vehtari ***@***.***> schrieb am Mi., 13. Nov. 2024, 18:45:

…

Using the github master I didn't get zombies with the minimal example, but after running at least 333 lines of brms_demo.R I get zombies when calling brms::log_lik() also with this PR — Reply to this email directly, view it on GitHub <#1704 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADCW2AGQTDBBIX7OFQ6KSFD2AOF3DAVCNFSM6AAAAABRWYXL4SVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDINZUGMZDCNJWGU> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[avehtari]

A new minimal example in the issue

[paul-buerkner]

Thanks. Okay, here is my next try. It contains two changes.

First, log_lik will no longer automatically use the mc.cores option, at least for now. Instead, it will only use parallel execution if cores is explicitely specified in log_lik. That is, your latest example

library(brms)
options(brms.backend = "cmdstanr", mc.cores = 2)

m0 <- brms::brm(yield ~ 0, npk, refresh=0)
ll<-brms::log_lik(m0) # no zombies

m1 <- brms::brm(yield ~ 1, npk, refresh=0)
ll<-brms::log_lik(m1) # yes zombies

will now almost surely not created zombies anymore simply because parallel execution of log_lik is not enabled via options(mc.cores = 2) anymore. This at least prevents the issue when log_lik is more or less accidentally executed in parallel.

However,

library(brms)
options(brms.backend = "cmdstanr", mc.cores = 2)

m0 <- brms::brm(yield ~ 0, npk, refresh=0, cores = 2)
ll<-brms::log_lik(m0) # no zombies

m1 <- brms::brm(yield ~ 1, npk, refresh=0, cores = 2)
ll<-brms::log_lik(m1) # yes zombies

may still created zombies.

In an attempt to "solve" this as well, I now force the user of PSOCK clusters in plapply. Perhaps when using PSOCK clusters, zombies will no longer occur? This is what I would like you to try out next with the latest version of the log-lik-zombies branch.

[n-kall]

I confirm that with master and previous fix, @avehtari's new minimal example was producing zombies.
Now with the new patch, zombies are not created even if I add cores = 2 or cores = 16 to the brm call or the log_lik call, and I do see additional R processes spawning as expected

[n-kall]

Edit: Aside note, I accidentally passed the mc.cores and cores args to brm call and with rstan I get the following uninformative error. With cmdstanr backend I get no error

devtools::load_all()
options(brms.backend = "rstan")

m0 <- brms::brm(yield ~ 0, npk, refresh = 0)
ll<-brms::log_lik(m0) # no zombies

m1 <- brms::brm(yield ~ 1, npk, refresh = 0, mc.cores = 4, cores = 4) # here mc.cores is not an accepted arg
#> Error in x@mode :
#>   no applicable method for `@` applied to an object of class "try-error"
#> In addition: Warning message:
#> In parallel::mclapply(1:chains, FUN = callFun, mc.preschedule = FALSE,  :
#>   4 function calls resulted in an error

[paul-buerkner]

you seem to be passing cores twice Noa Kallioinen ***@***.***> schrieb am Do., 14. Nov. 2024, 08:15:

…

Using cores argument seems to cause an error with the rstan backend though: devtools::load_all() options(brms.backend = "rstan") m0 <- brms::brm(yield ~ 0, npk, refresh = 0)ll<-brms::log_lik(m0) # no zombies m1 <- brms::brm(yield ~ 1, npk, refresh = 0, mc.cores = 4, cores = 4)#> Error in ***@***.*** :#> no applicable method for `@` applied to an object of class "try-error"#> In addition: Warning message:#> In parallel::mclapply(1:chains, FUN = callFun, mc.preschedule = FALSE, :#> 4 function calls resulted in an error — Reply to this email directly, view it on GitHub <#1704 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADCW2ABAHZBFMDKDCKXQ5FT2AREYDAVCNFSM6AAAAABRWYXL4SVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDINZVGU4DMMJZGE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[avehtari]

No more zombies (cmdstanr/rstan) with

m0 <- brms::brm(yield ~ 0, npk, refresh=0)
ll<-brms::log_lik(m0, cores=4) 

m1 <- brms::brm(yield ~ 1, npk, refresh=0)
ll<-brms::log_lik(m1, cores=4) 

I can also see 4 cores being used.

Also brms_demo.R does not produce zombies, but I guess it is now using just one core for log_lik computations inside powerscaling

[paul-buerkner]

Lovely, thank you both! I will clean up the PR and then merge :-)

[paul-buerkner]

Merged now :-)