Glossary

# Glossary

One line per term, with the module that teaches it. Grouped by topic.

---

## Jobs and Scheduling

| Term | Meaning | Learn more |
|------|---------|------------|
| **Allocation** | The nodes, CPUs, memory and GRES granted to a job for its lifetime. `sbatch`, `salloc` and `srun` all create one | [Submitting Jobs](../user/submitting-jobs.md) |
| **Step** | A unit of work launched with `srun` inside an allocation (`<jobid>.0`, `.1`, ...). `sacct` lists steps separately from the job | [Monitoring Jobs](../user/monitoring-jobs.md) |
| **Task** | One process, usually one MPI rank. `--ntasks` counts tasks; `--cpus-per-task` gives each task threads | [Resource Requests](../user/resource-requests.md) |
| **Job array** | Many near-identical jobs submitted at once with `--array`; each element has `SLURM_ARRAY_TASK_ID` | [Job Arrays](../user/job-arrays.md) |
| **Dependency** | A rule that holds a job until another job finishes (`--dependency=afterok:<id>`); reason `Dependency` | [Job Dependencies](../user/job-dependencies.md) |
| **Backfill** | A scheduler pass that starts lower-priority jobs now if they cannot delay higher-priority ones. Short, accurate `--time` fits more gaps | [Policies and Priority](../admin/policies-priority.md) |
| **Requeue** | Putting a job back in the queue after preemption, node failure or `scontrol requeue` | [Managing Jobs](../user/managing-jobs.md) |
| **Preemption** | A higher-priority job takes resources from a running one, which is requeued, suspended or cancelled | [Partitions and QOS](../admin/partitions-qos.md) |
| **Reservation** | Nodes set aside for a user, account or time window; jobs wait with reason `Reservation` until it opens | [Policies and Priority](../admin/policies-priority.md) |
| **scrontab** | Slurm's cron: recurring jobs defined with `scrontab -e` and run as ordinary jobs | [Recurring Jobs](../user/recurring-jobs-scrontab.md) |
| **Reason code** | The `REASON` column in `squeue` explaining why a job is pending | [Why Pending, Why Failed](why-pending-why-failed.md) |

---

## Accounting and Limits

| Term | Meaning | Learn more |
|------|---------|------------|
| **Account** | A group that jobs are charged to (`--account`); accounts nest in a tree | [Accounts and Fairshare](../admin/accounts-fairshare.md) |
| **Association** | The unit that limits and usage attach to: cluster + account + user, optionally + partition | [Accounts and Fairshare](../admin/accounts-fairshare.md) |
| **QOS** | Quality of Service: a named set of limits, priority and preemption rules a job can request or be assigned | [Partitions and QOS](../admin/partitions-qos.md) |
| **Fairshare** | A priority factor that lowers for accounts and users who have used more than their share recently | [Accounts and Fairshare](../admin/accounts-fairshare.md) |
| **TRES** | Trackable resources: `cpu`, `mem`, `node`, `billing`, `gres/gpu`, licenses. Limits and accounting are expressed in TRES | [Monitoring and Accounting](../admin/monitoring-accounting.md) |
| **Priority** | The number that orders pending jobs, built from age, fairshare, QOS and size factors (`sprio`) | [Policies and Priority](../admin/policies-priority.md) |
| **slurmdbd** | The accounting daemon that stores job history, associations and limits in a database | [Slurm Architecture](../admin/slurm-architecture.md) |
| **sacct / sreport** | `sacct` shows per-job history; `sreport` aggregates usage by user, account or TRES | [Monitoring and Accounting](../admin/monitoring-accounting.md) |
| **ExitCode** | `return_code:signal` in `sacct`. Read `State` first | [Job State Codes](job-state-codes.md) |
| **License (Slurm)** | A counted resource for software licenses, requested with `-L`; tracked locally or remotely | [Resource Management](../admin/resource-management.md) |

---

## Resources and Nodes

| Term | Meaning | Learn more |
|------|---------|------------|
| **Partition** | A named group of nodes with its own limits (time, size, access). Called a "queue" on ParallelCluster and PCS | [Resource Requests](../user/resource-requests.md) |
| **Node states** | `idle`, `mixed`, `alloc`, `down`, `drain`, and suffixes such as `~` (powered down) | [Node State Codes](node-state-codes.md) |
| **GRES** | Generic resources, mainly GPUs, requested with `--gres=gpu:N` | [GPU Jobs](../user/gpu-jobs.md) |
| **MIG** | Multi-Instance GPU: one physical NVIDIA GPU split into isolated slices; a slice is not a whole card | [GPU Jobs](../user/gpu-jobs.md) |
| **Shard** | `gres/shard`: several jobs share one GPU, scheduling only, processes are not fenced | [GPU Jobs](../user/gpu-jobs.md) |
| **MPS** | `gres/mps`: CUDA Multi-Process Service; a job asks for a percentage of one GPU | [GPU Jobs](../user/gpu-jobs.md) |
| **cgroup** | Linux kernel control group. Slurm uses it to enforce memory, CPU and device limits per job | [Resource Management](../admin/resource-management.md) |
| **Exclusive** | `--exclusive`: the job gets whole nodes and shares them with no one | [Resource Requests](../user/resource-requests.md) |
| **Constraint / feature** | A node tag requested with `--constraint` to pick hardware | [Resource Requests](../user/resource-requests.md) |
| **Environment module** | A named software environment loaded with `module load` (Lmod or Environment Modules) | [Environment Modules](../user/environment-modules.md) |

---

## Cluster Architecture

| Term | Meaning | Learn more |
|------|---------|------------|
| **slurmctld** | The controller daemon: schedules jobs and tracks nodes | [Slurm Architecture](../admin/slurm-architecture.md) |
| **slurmd** | The per-node daemon that launches and supervises job steps | [Slurm Architecture](../admin/slurm-architecture.md) |
| **sackd** | The authentication daemon that lets a login node run `sbatch`/`squeue` without running `slurmd` | [Slurm Architecture](../admin/slurm-architecture.md) |
| **Configless** | `slurmd` and `sackd` fetch their configuration from `slurmctld` instead of a local `slurm.conf` | [Slurm Architecture](../admin/slurm-architecture.md) |
| **slurmrestd** | The REST API daemon for submitting and querying jobs over HTTP | [Slurm Architecture](../admin/slurm-architecture.md) |
| **MUNGE** | The authentication service between Slurm daemons and clients | [Installation](../admin/installation.md) |
| **Login node** | The shared node users log in to for editing, submitting and light work; never for computation | [Best Practices](../user/best-practices.md) |
| **Head node** | ParallelCluster's name for the node running `slurmctld` (and `slurmdbd` if enabled), which users also log in to | [AWS ParallelCluster](../deployment/aws-parallelcluster.md) |
| **Scratch vs shared** | Scratch is fast storage for running jobs, often node-local; shared storage is visible from every node. Stage in and out | [Best Practices](../user/best-practices.md) |
| **SLUID** | **26.05+:** a Slurm-wide unique job identifier, exported as `SLURM_JOB_SLUID`; accepted by `squeue -j` | [Monitoring Jobs](../user/monitoring-jobs.md) |

---

## Cloud and Platform

| Term | Meaning | Learn more |
|------|---------|------------|
| **State=CLOUD** | A node that exists only when powered up; Slurm creates and destroys the instance behind it | [AWS ParallelCluster](../deployment/aws-parallelcluster.md) |
| **Power save** | `SuspendProgram`/`ResumeProgram`: Slurm powers idle cloud nodes down and starts them when jobs need them | [AWS ParallelCluster](../deployment/aws-parallelcluster.md) |
| **Protected mode** | ParallelCluster disables a queue (partition `INACTIVE`, fleet `PROTECTED`) after repeated consecutive node bootstrap failures (default 10) | [AWS Troubleshooting](../deployment/aws-troubleshooting.md) |
| **Node group** | AWS PCS's unit of compute: a set of instances of one configuration, mapped to a Slurm partition via a queue | [AWS PCS](../deployment/aws-pcs.md) |
| **Queue (cloud)** | ParallelCluster and PCS name for a partition; its compute resources define instance types and counts | [AWS ParallelCluster](../deployment/aws-parallelcluster.md) |
| **Static vs dynamic node** | Static nodes always run; dynamic nodes are launched on demand and released when idle | [AWS ParallelCluster](../deployment/aws-parallelcluster.md) |

Glossary

References¶