Reference

Glossary

  1. HOME
  2. Documentation
  3. Reference
  4. Glossary

Draft

This documentation is in draft and under active review. Figures for the Fall 2026 GPU reservation system are not yet published, and some pages describe behavior that has not been verified against a primary source. Confirm any figure with parasol@ucsd.edu before relying on it.

Look up a term used in the Datahub and PARASOL (formerly DSMLP) documentation. Entries are in alphabetical order, and each links to the page that covers the term in full.

A-C

Term Definition Reference
Active Directory (AD) The campus credential set, whose username and password, with Duo, authenticate to both Datahub and the login node. Datahub in the Browser
Anchor mode The rule that fixes a Service Unit budget window to the calendar, setting the point from which the window is measured. Anchor Modes
Auditor A read-only privilege tier held by staff. Managing a Group
Background pod (-b) A pod created and left running while the shell returns to the login node. Background Pods
Batch job (-B) A job run to completion unattended. Batch Jobs
Best-effort reservation A session admitted with no runtime guarantee and no charge. Not enabled on PARASOL. Best-Effort Reservations
Borrowing The use of idle GPU capacity beyond a workspace's quota. Borrowing Beyond Quota
Budget window The period a Service Unit budget covers before it renews. Budget Windows & Cadences
Claim window The period at the start of a booked window by which a matching session must be waiting or running, or the booking is cancelled as a no-show. The Claim Window
Cluster status page The page at datahub.ucsd.edu/hub/status that lists the GPU models on each node and how many are free. The Status Page
Cohort A group of workspaces whose quotas may sum to more than the physical capacity behind them, so that availability can read zero while a workspace has headroom. Cohorts
Container The isolated environment one job runs in, built from an image and run under the member's own UID without root. Root Access and System Packages
Course workspace A workspace covering one course: its lecture timeslots and the discussion sections associated with them. What a Workspace Is and What It Controls
Cull Short form of idle culling. What Counts as Idle

D-G

Term Definition Reference
Datahub The browser interface to PARASOL at datahub.ucsd.edu, comprising JupyterHub, the environment menu, and the services attached to it. Datahub in the Browser
datahub-base-notebook A former base image, no longer built or updated. Derive from datascience-notebook instead. Legacy Image Names
datascience-notebook The standard CPU image, containing widely used data analysis libraries from the Python, R, and Julia communities, and the base of the other standard images. Image Inheritance
/datasets The cluster-wide directory tree where shared corpora are staged, to be read in place rather than copied. Datasets
DeadlineExceeded The pod status reported when a container reaches the runtime limit, which does not indicate an error in the code. The DeadlineExceeded Status
disk-quota-service The service under the Services tab at datahub.ucsd.edu/hub/spawn that reports usage against quota and displays the workspace ID. Workspace and Personal Quotas
dsmlp-login.ucsd.edu The hostname of the login node. Connecting over SSH
Duo The campus two-factor prompt, which applies to both web and SSH access. Duo Authentication
Eviction The deletion of a running session because its reservation was cancelled or given to another user, with no warning beforehand. Eviction
Extend The reservation app's action that adds time to a running job's guarantee: it starts a new booking now, ending later than the job's current reservation, and moves the job onto it; called "continue" in the reservation system's interface. Extend
GHCR The GitHub Container Registry, where the standard images are published as ghcr.io/ucsd-ets/<image>:<tag>. Standard Images, Tags, and Pinning
git-pull link A URL that fetches a repository into an environment, commonly used to distribute course materials. Sign-In & Session Problems
Globus A transfer service for moving large volumes of data in and out of the platform. Moving & Sharing Data
GPU class A GPU size band, rather than a hardware model, requested with -l gpu-class=<class>; each workspace is granted access to one or more classes. GPU Classes
Grader account The single shared account each course receives for nbgrader and formgrader, which, unlike an instructor's own account, holds write permission on the workspace's public/ directory. Grader Account
Group manager Another name for workspace manager. Managing a Group

H-L

Term Definition Reference
Home directory Per-user, per-workspace storage, of which a member holds a separate one in every workspace they belong to. Workspace and Personal Quotas
Idle culling The reclaiming of a GPU from a session that has stopped using it. What Counts as Idle
Image The packaged filesystem a container starts from. Standard Images, Tags, and Pinning
Jumpbox Another name for the login node: a machine for launching jobs and moving files, on which running computation is prohibited. What the Login Node Is For
JupyterHub / JupyterLab The hub that spawns a session, and the notebook interface it spawns; replacing lab with tree in the URL reaches the older Notebook interface. Datahub in the Browser
Kubernetes The orchestration system on which the platform runs, in which every job is a pod in a per-user namespace. Direct Kubernetes Use and Session Events
kubesh <pod-id> The command that enters one of an account's own running pods from the login node. A Second Shell with kubesh
launch.sh The launcher at /opt/launch-sh/bin/launch.sh, which maps command-line parameters into a Kubernetes pod definition. launch.sh Reference
Launch wrapper A script such as launch-scipy-ml.sh or launch-datascience.sh that sets environment variables and then hands off to launch.sh. Default Resources
Limit and request The two values Kubernetes holds for CPU and memory: the limit is the value passed at launch, and the request is the amount actually reserved. Resource Requests and Limits
Login node The host dsmlp-login.ucsd.edu, reached by SSH with AD credentials and Duo, whose home directory ~ is the personal area rather than a course home. Connecting over SSH

M-P

Term Definition Reference
Manual resetter A service under the services dropdown at datahub.ucsd.edu that stops an account's servers, signs it out, and resets its profile while preserving files. "Spawn Failed"
Member The ordinary privilege tier, held by a student in a course or a member of a lab. Managing a Group
MIG slice An isolated share of one GPU card's memory and compute, created with NVIDIA Multi-Instance GPU (MIG). A session on a slice can use only that slice. MIG Slices
Namespace The per-user Kubernetes namespace that pods run in, and the unit against which one of the resource tiers applies across everything running at once. Resource Tiers
nbgrader The in-platform tool for distributing, collecting, and grading assignments, from which grade export to Canvas is manual. Grading
No-show A booking cancelled and charged because no matching session was waiting or running when its claim window closed, or soon after the last session under it ended. The Claim Window
Off-peak The discounted hours outside the peak evening period. Peak & Off-Peak Hours
On-demand lease The reservation the system creates when a GPU session is launched without a booking, which draws on the Service Unit budget as a booked window does. On-Demand Lease Charges
ORG_ON_DEMAND The default workspace, for on-demand leases only, that a GPU session naming no course is charged to, such as a launch without -W. The Default Workspace
OOMKilled The pod status reported when a container reaches its memory limit. Resource Requests and Limits
Otter-Grader, Gradescope Grading tools used on the platform alongside nbgrader. Choosing a Grading Tool & Interface
Overstay Running a session past the end of its runtime guarantee, unprotected but not charged. Overstay
PARASOL The Platform for Applied Research, Analytics, Science and Open Learning, the on-premises cluster of CPU and GPU nodes on which Datahub sessions and command-line jobs run, for courses, student projects and research alike, formerly called DSMLP. How PARASOL Works
Pinning Fixing a workspace to a dated image tag so that its members are not moved by an image update mid-term. Pinning a Workspace
Pod One running container as Kubernetes represents it; kubectl get pods lists the pods in an account's own namespace. Direct Kubernetes Use and Session Events
Pod label (-l key=value) A repeatable attribute attached to a pod at launch, and the means by which gpu-class is set. Resource and GPU Selection Flags
Preemption The ending of a session past its runtime guarantee because its GPU is needed, for a booking or to keep a share of the class free. Preemption
private/ A per-user, cluster-wide directory that appears unchanged in every workspace and draws on the personal quota rather than any workspace's. Where Files Live
public/ The workspace's shared area, readable by every member, written by the grader account in a course, and drawing on the workspace quota. Where Files Live
P3 / P4 University of California data protection levels: P4 data is prohibited on the platform, and P3 data may be permitted after review. Data Classification

Q-S

Term Definition Reference
Quota, GPU The ceiling on how many GPUs of each class a workspace may hold at one time, separate from a Service Unit budget. What a Quota Is
Quota, storage Two separate pools: the workspace pool, covering the workspace home, public/, and teams/, and the personal pool, covering private/ and following the member into every workspace. Workspace and Personal Quotas
Reservation A booked window during which GPU capacity is held and a session is admitted ahead of the walk-up queue; the session is still launched as usual. Reservations
Reservation app The web app at reserve.dsmlp.ucsd.edu where GPU windows are booked, cancelled, and extended. The Reservation App
Reservation event A Kubernetes event the reservation system writes on a GPU pod, such as OnDemandLeaseDenied or Preempted. Reservation Events
Reserve floor Capacity that ITS keeps unborrowable in each GPU class, so that idle capacity is not always available. The Reserve Floor
Resource tiers The three levels at which resource limits apply to a job: a single pod, the namespace in total, and what is available on request. Resource Tiers
rstudio-notebook The standard image that adds the RStudio environment to datascience-notebook; it derives from the CPU image and is not GPU-enabled. Standard Images
Runtime guarantee The period in which a session cannot be preempted: to the end of its reservation, and through any booking that directly follows it. End of a Reservation Window
Runtime limit The deadline on how long a container may run. The Runtime Limit
scipy-ml-notebook The standard image that adds CUDA/GPU tooling, TensorFlow, and PyTorch to datascience-notebook, and the default image for bare launch.sh. Standard Images
Senior and junior borrowing The two borrowing tiers. A junior borrower may borrow only the idle capacity above a deeper reserve floor; the tier applies only when a reservation is made. Borrowing Beyond Quota
Service Unit (SU) A usage credit for GPU time, spent at an hourly rate that each GPU class carries. What a Service Unit Is
Slurm compatibility wrappers The commands sbatch, srun, squeue, and scancel, which translate into launch.sh and have no Slurm scheduler behind them. Slurm Compatibility Wrappers
Spawn Starting a browser session from a course's environment menu; a spawn failure is the message returned when the session does not start, and it does not state the cause. "Spawn Failed"
Standard image One of the images ITS maintains, which receive priority support. Standard Images

T-Z

Term Definition Reference
Team A group within a workspace, such as a project group, a lab bench, or a capstone team, holding data scoped to some of its members. Sections, Teams and Group Data
Team mode A reservation feature in which a team shares one Service Unit budget, and teammates can see, cancel, and adopt one another's bookings. Team Mode
teams/ The directory where team-scoped data appears, drawing on the workspace quota. Where Files Live
TPOC (Technical Point of Contact) The person a course nominates to lead its customization work, a designation for support purposes rather than a privilege tier. Managing a Group
TritonLink / TSS The registration systems from which course rosters populate automatically; auditors and observers are not on them and are added through Canvas. Students Enrolled in a Course
Universal Scale Storage (USS) SDSC storage that can be mounted onto the cluster, supported by Research IT. Mounting External Storage
Unprivileged container A container running under the member's own UID, without root or sudo, in a per-user Kubernetes namespace. Root Access and System Packages
Workspace The unit for one course, lab, or catch-all population, which anchors its roster, storage, container images, GPU class access, quotas, and budgets. What a Workspace Is and What It Controls
Workspace manager The privilege tier held by an instructor, a TA, or a PI. Managing a Group

Former Names

DSMLP, the Data Science & Machine Learning Platform, is the former name of PARASOL. It is still part of hostnames and system output, such as dsmlp-login.ucsd.edu.

Research Cluster was the name for research work on what is now PARASOL. Support for research work is from Research IT at rcd-support@ucsd.edu.

Report a problem with this page