Using sjp in job submit.lua - KasperSkytte/slurmjobpacker GitHub Wiki

Using sjp in job_submit.lua

Requires JobSubmitPlugins=lua in slurm.conf.

local ok, sjp = pcall(dofile, "/opt/slurmjobpacker/lua/sjp.lua")
if not ok then
    slurm.log_error("job_submit: cannot load sjp.lua: %s", tostring(sjp))
    sjp = nil
end

function slurm_job_submit(job_desc, part_list, submit_uid)
    -- your own rules first

    if sjp and sjp.place(job_desc, submit_uid) then
        return slurm.SUCCESS
    end

    -- not placed: your fallback, e.g. by memory per CPU
    local cpus = job_desc.min_cpus or 1
    local per_cpu = job_desc.min_mem_per_cpu or (job_desc.min_mem_per_node or 0) / cpus
    job_desc.partition = per_cpu >= 6000 and "fat1,fat2" or "slim1,slim2"
    return slurm.SUCCESS
end

sjp.place() never rejects a job. It returns true, or false and a reason, leaving the job to your rules (only its AdminComment says why: sjp:skipped=<reason>, if empty):

reason
"no room" no node that could hold the job has room now
"no table" sjpd not running, or disable file present
"stale" sjpd data too old
"not batch" salloc/srun
"gpu" GPUs requested
"reservation" in a reservation
"constraint" --constraint: the user asks for particular hardware
"no usable node in --nodelist" the --nodelist nodes are down or not in sjp's partitions
"no partition holds those nodes" no partition holds every --nodelist node
"multi-node" -N 2 or more
"no memory" --mem=0
"error" see the slurmctld log 
local placed, why = sjp.place(job_desc, submit_uid)

When placed: sets partition, req_nodes (pins) and admin_comment (if empty). A job's own partition is replaced. Its --nodelist and --exclude are kept and sjp chooses among those nodes, or with [policy] user_nodes = "ignore" they are dropped.

Settings

default
sjp.config.verbose false detailed log line per job
sjp.config.table_path /run/sjp/policy.lua if [general] state_dir moved
sjp.config.disable_path /etc/sjp/disable if [general] disable_file moved

Example script

lua/job_submit.lua: salloc/srun to interactive, GPU jobs to GPU_PARTITION, the rest to sjp, with an optional SLIM/FAT/FAT_ABOVE fallback. The Ansible role can install it (sjp_job_submit_install: true).

After upgrading sjp: scontrol reconfigure (Slurm doesn't reload dofiled files).

⚠️ **GitHub.com Fallback** ⚠️