How To See The Sbatch Command That Gets Run

Get Help
1

I’m trying to apply a GPU with Slurm using native attributes. It runs but the GPUs aren’t being reserved as expected. How can I see the actual sbatch command that is running to know what is going on? My actual submit.yml.erb file is given below:

batch_connect:
  template: "basic"
<%-
  slurm_args = if gpu_switch == 1
                 ["--gpus-per-node", "1", "--gres", "gpu:1" ]
               else
                 []
               end
-%>

script:
  native:
  <%- slurm_args.each do |arg| %>
    - "<%= arg %>"
  <%- end %>
2

My guess is it’s a type issue where gpu_switch is a string not an integer.

Either turn gpu switch into a integer with gpu_switch.to_i or compare it against the string 1 like == "1".

Also, just as an aside to style, you may want to put the ERB <%- %> block at the top for readability.

<%- 
   # do all the computations up here
%>
---
batch_connect:
  # and so on
3

Yep, that was it… thanks!

Is there a way to see the sbatch command that is actually getting run? And any other tips for debugging these kinds of issues?

4

Yea you can see the command being run in /var/log/ondemand-nginx/USER/error.log. For this error, I just kinda knew from experience that they’re always strings - maybe I’d run into the same issue once or twice. Something like this, gpu_switch isn’t going to be in the sbatch command, but you can see what all these templates got templated with in the user_defined_context.json in the job’s session directory (the link in the card). This will indicate whether or not flags are being correctly passed from the form to the submit.yml.

5

I’m not seeing much useful in error.log. The user_defined_context.json is definitely helpful.
If I drop print statements into my ERB block in submit.yml.erb, I cannot find the output anywhere. Is there a way to print out info like this?

6

No, but if you’re in a development environment - that is, in your home directory where only you have access to this app - I have raised errors with debug messages like:

<%-
  raise StandardError.new("the value of the things I'm looking for is: #{the_thing.inspect}")
-%>
7

Not sure how useful this is, but:

% cat submit.yml.erb                      
# Job submission configuration file
#
---

<%-
slurm_args = if gpu_switch.to_i == 1
                 #["--nodes", "#{nodes}", "--ntasks-per-node", "#{ppn}", "--gpus-per-node", "1", "--gres", "vis" ]
                 ["--gpus-per-node", "1", "--gres", "gpu:1" ]
               else
                 []
               end
-%>

batch_connect:
  template: "basic"

script:
  native:
  <%- slurm_args.each do |arg| %>
    - "<%= arg %>"
  <%- end %>
   - "--exclusive"%

Manually process the template using CLI erb, providing params:

% erb  -T - gpu_switch="1" submit.yml.erb 

# Job submission configuration file
#
---


batch_connect:
  template: "basic"

script:
  native:

    - "--gpus-per-node"

    - "1"

    - "--gres"

    - "gpu:1"

   - "--exclusive"
8

Sure is! If you end the each and end lines with -%> instead of %> you’ll get rid of the spaces in the array arguments.

Also the last - "--exclusive" seems to be off by 1 space.