Glossary

Common HPC terms and their definitions.

A

Allocation: A grant of computing resources (CPU hours, storage) to a research group.
Apptainer: Container runtime for HPC (formerly Singularity). Allows running containerized applications without root privileges.

Batch Job: A job submitted to the scheduler to run without user interaction. Contrast with interactive job.

Cluster: A collection of interconnected computers (nodes) that work together as a single system.
Compute Node: A server in the cluster dedicated to running user jobs. Users cannot log in directly; jobs are assigned by the scheduler.
Container: A lightweight, standalone package containing code and dependencies. See Apptainer, Docker.
Core: A single processing unit within a CPU. Modern CPUs have multiple cores (e.g., 64 cores per CPU).
CPU (Central Processing Unit): The main processor in a computer. HPC nodes typically have 1-2 CPUs with many cores each.

Docker: Popular container platform. Docker images can be converted to Apptainer format for HPC use.

Environment Module: System for managing software packages. Users load modules to access specific software versions.

Fairshare: Scheduling policy that balances resource allocation based on historical usage. Groups that recently used many resources get lower priority.
FLOPS: Floating-point Operations Per Second. Measure of computing performance.

Globus: High-performance file transfer service for moving large datasets between systems.
GPU (Graphics Processing Unit): Specialized processor for parallel computations. Essential for deep learning and certain scientific simulations.

Head Node: Server that manages the cluster, runs the scheduler, and routes network traffic.
HPC (High Performance Computing): Using supercomputers and clusters to solve complex computational problems.
HTC (High Throughput Computing): Running many independent jobs, optimizing for total work completed rather than individual job speed.

InfiniBand: High-speed, low-latency network interconnect used between cluster nodes. Much faster than Ethernet.
Interactive Job: A job where the user directly interacts with the compute node via terminal. Contrast with batch job.

Job: A unit of work submitted to the cluster scheduler.
Job Array: A collection of similar jobs submitted as a single entity, each with a unique index.
Job Script: A shell script containing SLURM directives and commands to execute.

Login Node: Server where users connect via SSH. Used for file editing, job submission, and light tasks. Not for heavy computation.

MFA (Multi-Factor Authentication): Security requiring multiple verification methods (password + Duo app).
Module: See Environment Module.
MPI (Message Passing Interface): Standard for parallel programming across multiple nodes. Programs use MPI to communicate between processes.

Node: A single computer/server in the cluster.
NUMA (Non-Uniform Memory Access): Memory architecture where access time depends on memory location relative to the processor.

OOD (Open OnDemand): Web-based interface for accessing HPC resources via browser.
OpenMP: API for shared-memory parallel programming. Programs use OpenMP for multi-threaded parallelism within a single node.

Parallel Computing: Running computations simultaneously across multiple cores, GPUs, or nodes.
Partition: A logical grouping of nodes in SLURM (e.g., gpu partition for GPU nodes).
PI (Principal Investigator): Faculty member leading a research group. PIs authorize group membership and manage allocations.
PTA: Project-Task-Award. Caltech’s accounting code for charging compute usage.

QOS (Quality of Service): SLURM setting that modifies job priority or limits. Example: debug QOS for quick test jobs.
Queue: List of jobs waiting to run. Jobs are selected based on priority and resource availability.
Quota: Storage limit (e.g., 50 GB home directory quota).

SBATCH: SLURM command to submit a batch job.
Scheduler: Software that manages job queue and allocates resources. Caltech uses SLURM.
Scratch: Temporary high-performance storage for active computations. Files are automatically deleted after 14 days.
SIF (Singularity Image Format): Container image format used by Apptainer/Singularity.
SLURM (Simple Linux Utility for Resource Management): The job scheduler and resource manager used on the cluster.
SSH (Secure Shell): Protocol for secure remote login and command execution.
SRUN: SLURM command to run parallel tasks within a job.

Task: A single process in a SLURM job. A job can have multiple tasks.
Thread: A lightweight execution unit within a process. Multi-threaded programs run multiple threads per core.

VAST: High-performance storage system. Current primary storage at Caltech HPC.
VPN (Virtual Private Network): Secure connection to campus network from off-campus locations.

Walltime: Maximum allowed runtime for a job. Jobs exceeding walltime are terminated.
Worker Node: See Compute Node.

Acronym	Meaning
CPU	Central Processing Unit
GPU	Graphics Processing Unit
HPC	High Performance Computing
HTC	High Throughput Computing
MFA	Multi-Factor Authentication
MPI	Message Passing Interface
OOD	Open OnDemand
PI	Principal Investigator
QOS	Quality of Service
SIF	Singularity Image Format
SLURM	Simple Linux Utility for Resource Management
SSH	Secure Shell
VPN	Virtual Private Network