README.md

<!-- README.md is generated from README.Rmd. Please edit that file -->

# Connect to GitLab

If you haven’t used git at the MPI yet, you should generate SSH keys and
upload the public key to Gitlab (see the [Gitlab
docs](https://docs.gitlab.com/ee/gitlab-basics/create-your-ssh-keys.html)).

(Skip if you have ssh keys): Open a terminal and run:

``` bash
ssh-keygen -t ed25519 -C "peikert@mpib-berlin.mpg.de"
```

to create the new keys. Open the newly generated public key (or your
existing one) and copy it. Put it here
<https://git.mpib-berlin.mpg.de/profile/keys>.

# Connect to Tardis

For the cluster you have to setup two additional protocols, namely VPN,
SSH for public Git projects. For private GitLab projects you need yet
another protocol (GitLab Tokens).

## Connect to VPN

VPN enables you to MPI internal resources (aka ip addresses) like
printers or TARDIS, the beloved cluster.

Follow the steps of the wiki:

<https://wiki.mpib-berlin.mpg.de/books/it/page/vpn>

## Connect to Tardis with SSH

To connect to tardis you have copy your ssh keys to it so it recognizes
you without password.

### Using ssh-copy-id

The `ssh-copy-id` command does this for you:

``` bash
ssh-copy-id peikert@tardis.mpib-berlin.mpg.de
```

Test connection:

``` bash
ssh peikert@tardis.mpib-berlin.mpg.de
```

If you can login without password everything works. If not ask Michael
Krause.

### If ssh-copy-id does not work

But `ssh-copy-id` is not available everywhere.

You can also login to tardis and paste the public key (usually
`~/.ssh/id_rsa.pub`) as a newline in the file `~/.ssh/authorized_keys`
(if it is empty does not exist, don’t worry just create it):

``` bash
ssh peikert@tardis.mpib-berlin.mpg.de
# this command edits/creates the file
# paste content of ~/.ssh/id_rsa.pub on its own line
nano ~/.ssh/authorized_keys
```

Test connection:

``` bash
ssh peikert@tardis.mpib-berlin.mpg.de
```

If you can login without password everything works. If not ask Michael
Krause.

## Your first future, brought to you by tardis

Remember to activate VPN. You should now be able to connect to tardis:

``` r
tardis <- parallelly::makeClusterPSOCK("tardis.mpib-berlin.mpg.de",
                                       port='random', user="peikert",
                                       # this R version does not matter
                                       # it only requires that future.batchtools is there
                                       rscript=c("/opt/software/R/4.0.3/bin/Rscript"), 
                                       homogeneous = TRUE)
```

And you can use futures to evaluate stuff on tardis.

``` r
library(future)
# set option to ignore a warning
options(future.rng.onMisuse = "ignore")
plan(tweak(cluster, workers=tardis))
value(future(4*3)) # 4*3 is calculated on tardis login node
#> [1] 12
```

Anyhow, while this works you do not use the full power of tardis because
these things are only calculated on the login node of tardis (and you
should never do this for computing intense tasks). You need to use slurm
to submit jobs and request resources.

``` r
library(future.batchtools)
plan(list(tweak(cluster, workers=tardis),
          tweak(batchtools_slurm,
                workers = 1,
                resources=list(ncpus=1,
                               memory='200m',
                               walltime=600,
                               partition=c('quick')))))
```

Then you use a second layer of futures to evaluate things on tardis’
workers:

``` r
local <- 2*4
login_node <- value(future(2*4))
worker <- value(future(value(future(2*4))))
```

It may be that this does not work for you, because the package and R
versions on tardis, its workers and your computer have to match. To
solve this we use Docker and Singularity.

# Align your R with Tardis

## Docker on GitLab

Docker provides a container with an operating system and software that
is portable across machines. Real handy if you want to ensure the same
software across machines. The great thing is you do not even need Docker
installed for this to work, because tardis just needs the container
image and you do not have to build it yourself. The image is build by
GitLab from the `Dockerfile` which you will create in a minute. To align
tardis with the software you use yourself, first update all your
packages then edit the [Dockerfile provided in this
repo](https://git.mpib-berlin.mpg.de/peikert/a-future-for-tardis/-/blob/master/Dockerfile).
Change the version number in line 1 to the R version you use, change the
date to today and add packages at will to the list. Commit the
Dockerfile to Git/GitLab. While you have a Dockerfile, GitLab does not
yet know that it should build the image for you. Just copy and commit
the [`.gitlab-ci.yml` file provided in this
repo](https://git.mpib-berlin.mpg.de/peikert/a-future-for-tardis/-/blob/master/.gitlab-ci.yml)
When you have pushed it you see that GitLab is working on the page of
your repo on the left hand side under CI/CD. When it is finished you see
the image under “Packages & Registries” -\> “Docker Registry”.

## Singularity on Tardis

If you have made the repository public things are a little bit easier
because you do not need to authenticate to GitLab from Tardis to
download the Docker image. Login on Tardis:

``` bash
ssh peikert@tardis.mpib-berlin.mpg.de
```

``` bash
# navigate to your project
cd project
# pull image from gitlab
singularity pull docker://registry.git.mpib-berlin.mpg.de/peikert/a-future-for-tardis:latest
```

Now batchtools has to know that you want to use singularity. For that
you copy
[`.batchtools.slurm.singularity.tmpl`](https://git.mpib-berlin.mpg.de/peikert/a-future-for-tardis/-/blob/master/.batchtools.slurm.singularity.tmpl)
to tardis. And change the path of the singularity image in this
template.

You have to change the plan again to include the slurm template:

``` r
plan(list(tweak(cluster, workers=tardis),
          tweak(batchtools_slurm,
                workers = 1, #number of tasks that may run parallel on tardis
                # the R/package version of the singularity image has to match
                # the local machine, therefore check Dockerfile for R version
                template = "/home/mpib/peikert/.batchtools.slurm.singularity.tmpl",
                resources=list(ncpus=1,
                               memory='200m',
                               walltime=600,
                               partition=c('quick')))))
```

``` r
local <- 2*4
login_node <- value(future(2*4))
worker <- value(future(value(future(2*4))))
```

For this workflow only local and worker need to be aligned. Notice that
the login node runs Debian but the worker run Ubuntu because the
container are based on Ubuntu.

``` r
session <- function()with(sessionInfo(), list(rversion = paste0(R.version$major, ".", R.version$minor),
                          os = running))
list(local = session(),
     login_node = value(future(session())),
     worker = value(future(value(future(session())))))
#> $local
#> $local$rversion
#> [1] "4.0.4"
#> 
#> $local$os
#> [1] "Solus 4.2 Fortitude"
#> 
#> 
#> $login_node
#> $login_node$rversion
#> [1] "4.0.3"
#> 
#> $login_node$os
#> [1] "Debian GNU/Linux 10 (buster)"
#> 
#> 
#> $worker
#> $worker$rversion
#> [1] "4.0.4"
#> 
#> $worker$os
#> [1] "Ubuntu 20.04.2 LTS"
```