Improve app responsiveness - use CRIU for the R process

[tbrittoborges]

https://support.openanalytics.eu/t/create-running-user-proxies-containers-in-advance/2521/2

[tbrittoborges]

Hi @HaraldWilhelmi,

can you test what we discussed earlier today?

The app leaves in the shiny server: (10.250.140.12).

The docker file is at /home/tbrittoborges/apps/nmd_transcriptome/Dockerfile

Thanks!

[HaraldWilhelmi]

The present version of the app is based on the the Docker image rocker/verse:4.2.1. That in turn uses an old Ubuntu version (focal), for which no criu is available. So as a first step we moved to a private repo (git@github.com:dieterich-lab/nmdtx.git) and try to upgrade to a much newer image/R version (4.3.1). That will give us a much newer Ubuntu version, for which we have at least a pre-compiled PPA version.

[HaraldWilhelmi]

Using this instructions to upgrade renv.lock: https://inbo.github.io/tutorials/tutorials/r_renv_update/

[HaraldWilhelmi]

Had to do manually: R -e 'renv::install("dzhang32/ggtranscript")'

[HaraldWilhelmi]

Seems that it is easier to stay with the old software versions. The key seems to be to get criu from a PPA repo:

add-apt-repository ppa:criu/ppa

[tbrittoborges]

See here for the app initialization profile. profile.html.zip

[HaraldWilhelmi]

Status from Monday: We have a working implementation with CRIU now - see https://github.com/dieterich-lab/nmdtx/tree/main/deploy. With naive approach to just preload 'library('nmdtx') we just get an improvement of 2 seconds (13 s -> 11 s).

[tbrittoborges]

Profiling the app within the container is not trivial. Here are my steps. As of now, the profiler is not outputting the report as expected.

Profiling

Connect to the shiny-new server:

ssh tbrittoborges@10.250.140.14
cd /home/tbrittoborges/nmdtx/deploy

Change rstart.R by adding the preload command, in the example the command is commented out:

# library("nmdtx")
while (! file.exists('/restore.marker')) {
    print ('waiting....')
    print('Library loaded')
    Sys.sleep(0.5)
}
system(paste('mkdir -p', tempdir()))
system('tail --follow=name /criu_dumps/rlogs.log > /criu_dumps/rlogs_container.log  &')
shiny_user  <- basename(readLines('/shiny_user.txt'));
shiny_groups  <- basename(readLines('/shiny_usergroups.txt'));
print(paste('User', shiny_user))
Sys.setenv('SHINYPROXY_USERNAME' = shiny_user, 'SHINYPROXY_USERGROUPS' = shiny_groups)
print(Sys.getenv('SHINYPROXY_USERNAME'))
print(Sys.getenv('SHINYPROXY_USERGROUPS'))
options('shiny.port'=3838,shiny.host='0.0.0.0')
p <- profvis::profvis({
    print(nmdtx::run_app())
})
htmlwidgets::saveWidget(p, "profile.html")

Build the image:

./build.sh

Run the container on the detached mode, so we can copy the profviz output:

docker run --privileged -it --entrypoint '/restore.sh' --detach nmdtx

Keep track of the init.

docker logs -f $(docker ps -q | head -n1)

Once the app is done, copy the profile file from the container to the server.

docker cp $(docker ps -q | head -n1):profile.html profile.html

Copy the file from a remote to your local machine, give it a descriptive name.

scp tbrittoborges@shinynew:/nmdtx/deploy/profile.html  profile_noChange.html

Once done, stop the container, and profile the next initialization strategy.

docker stop $(docker ps -q | head -n1)

Important if the following message repeats after ./build.sh, there is something wrong rstart.R:

ls: cannot access '/criu_dumps/dump.done': No such file or directory
Waiting for dump...

[HaraldWilhelmi]

About:

ls: cannot access '/criu_dumps/dump.done': No such file or directory
Waiting for dump...

That comes from the loop, which check if the dump is done. It is normal for the start, but should stop after 2 or 3 interations.

[tbrittoborges]

Yes, that is fine. The problem is at the profiling line:

p <- profvis::profvis({
    print(nmdtx::run_app())
})

The profvis function requires a stop signal, and I could not find a way to do that within the docker environment.

[tbrittoborges]

For documentation, I tried the solution for remote server, from here: https://rstudio.github.io/profvis/faq.html#can-i-profile-a-shiny-application-running-on-a-remote-server

But current the app has not stop, but I could add one.

Rprof(strftime(Sys.time(), "%Y-%m-%d-%H-%M-%S.Rprof"),
      interval = 0.01, line.profiling = TRUE,
      gc.profiling = TRUE, memory.profiling = TRUE)

onStop(function() {
  Rprof(NULL)
})