Terms used across this documentation set, alphabetically. Where a term has a page of its own, the definition here is the one-line form and the link is the authority.
Active Directory (AD) — the campus credential set. An AD username and password, with Duo, authenticate to both Datahub and the login node. → Datahub in the Browser
Anchor mode — how a Service Unit budget window is fixed to the calendar: what "weekly" is weekly from. Which modes are configured here is not settled. → Budget Windows
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. The session is not placed inside it, and exiting a pod
does not stop processes started in it. Distinct from &, which backgrounds a
process rather than a pod. → Interactive, Background & Batch Modes
Batch job (-B) — a job run to completion unattended. -B -- <command> is
the usual form. → Interactive, Background & Batch Modes
Best-effort reservation — a session that runs now, free of a wait, and accepts preemption from its first tick. It holds nothing and is protected by nothing. → Best-Effort Reservations
Borrowing — using idle GPU capacity beyond a workspace's quota, available to work starting under roughly 12 hours ahead. Course workspaces always borrow senior, meaning they are not the first asked to give capacity back. → Quotas, Cohorts & Availability
Budget window — the period a Service Unit budget covers before it renews. Weekly for courses; monthly or quarterly for research, where one is set at all. → Budget Windows
Claim window — the 15 minutes within which a session must be started on a booked reservation. Missed, the reservation is cancelled, the capacity returns to the pool, and the window is gone for its remainder. → The Claim Window
Cluster status page — datahub.ucsd.edu/hub/status,
which lists the GPU models on each node and how many are free.
→ The Status Page
Cohort — a group of workspaces whose quotas may deliberately sum to more than the physical capacity behind them, allowing a higher shared peak. A consequence: availability may read zero while a workspace has headroom on paper. → Cohorts
Container — the isolated environment one job runs in, built from an image and run under the member's own UID without root. → Standard Images
Course workspace — a workspace covering one course: its lecture timeslots together with the discussion sections associated with them. A discussion section does not get a workspace of its own. → What a Workspace Is
Cull — see Idle culling.
Datahub — datahub.ucsd.edu, the browser front
door: JupyterHub, the environment menu, and the services that hang off it.
→ Datahub in the Browser
datahub-base-notebook — the smallest standard image, sitting below
datascience-notebook. The one to derive from when build time matters.
datascience-notebook — the standard CPU image: widely-used data analysis
libraries from the Python, R and Julia communities.
/datasets — the cluster-wide tree where shared corpora are staged, read in
place rather than copied. → Shared Datasets
DeadlineExceeded — the pod status meaning the runtime limit was reached. Not
an error in the code. → The Runtime Limit
disk-quota-service — the service reporting usage against quota, under the
Services tab at
datahub.ucsd.edu/hub/spawn. It also
displays the workspace ID. → Directories, Quotas & Cleaning Up
DSMLP — the Data Science / Machine Learning Platform: the on-premises cluster of CPU and GPU nodes underneath everything in this documentation.
dsmlp-login.ucsd.edu — the login node. See Login node.
Duo — the campus two-factor prompt, which applies to both web and SSH access.
GHCR — the GitHub Container Registry, where the standard images are published
as ghcr.io/ucsd-ets/<image>:<tag>. The older ucsdets/<image> naming is also
in circulation. → Standard Images
git-pull link — a URL that fetches a repository into an environment,
commonly used to distribute course materials. Its commonest failure is being
clicked before the student has signed in.
→ Sign-In & Session Problems
Globus — a transfer service used for moving data in and out at volume. → Moving & Sharing Data
GPU class — a size band rather than a hardware model: extra-small, small,
medium, large, extra-large, requested with -l gpu-class=<class>. Each
workspace is granted access to one or more. → GPU Classes
Grader account — the single shared account each course receives for nbgrader
and formgrader. It holds write permission on the workspace's public/ directory;
an instructor's own account does not. → Grading
Group manager — see Workspace manager.
Home directory — per-user, per-workspace storage. A member holds a separate one in every workspace they belong to. Course homes run 5-10 GB; research homes are on the order of 100 GB. → Directories, Quotas & Cleaning Up
Idle culling — reclaiming a GPU from a session that has stopped using it: roughly 30 minutes of idleness on a busy cluster, up to 6 hours on a quiet one, never in a session's first 45 minutes. A warning precedes it, and a cull is not a crash. CPU-only sessions are not culled by this mechanism. → Idle Culling
Image — the packaged filesystem a container starts from. See Standard image.
Jumpbox — what the login node is: a machine for launching jobs and moving files, not for computing. → The Login Node
JupyterHub / JupyterLab — the hub that spawns a session, and the notebook
interface it spawns. Swapping lab for tree in the URL reaches the older
Notebook interface.
Kubernetes — the orchestration system underneath the platform. Every job is a pod in a per-user namespace. → Kubernetes
kubesh <pod-id> — the command that enters one of an account's own running
pods from the login node.
launch.sh — the launcher, at /opt/launch-sh/bin/launch.sh, which maps
command-line parameters into a Kubernetes pod definition. Bare, it gives 1 CPU /
1 GB / 0 GPU. → launch.sh Reference
Launch wrapper — launch-scipy-ml.sh, launch-datascience.sh and their
relatives: scripts that set environment variables and then hand off to
launch.sh. They give 2 CPU / 8 GB / 0 GPU.
Limit and request — the two numbers Kubernetes holds for CPU and memory. The
number passed is the limit; the request — what is actually reserved — is half of
it. GPUs are exempt. → launch.sh Reference
Login node — dsmlp-login.ucsd.edu, reached by SSH with AD credentials and
Duo, and requiring the VPN from off campus. It shares a filesystem with the
containers. ~ there is the personal area, not any course home.
→ Connecting over SSH
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, preserving files. The remedy for a stale profile or a broken package in
an account's own .local. → Sign-In & Session Problems
Member — the ordinary privilege tier: a student in a course, a member of a lab. → Managing a Group
Namespace — the per-user Kubernetes namespace pods run in, and the unit one of the three resource tiers is expressed against — 8 CPU / 64 GB / 1 GPU by default, across everything running at once.
nbgrader — the in-platform assignment distribution, collection and grading tool. Grade export to Canvas is manual. → Grading
NoSchedule taint — the Kubernetes mechanism that keeps pods off a node
unless they carry a matching toleration. medium and above carry them, and a
GPU request with no gpu-class label fails with 0/5 nodes available.
→ Error Messages
Off-peak — the discounted hours, which steer usage away from peak evening demand. → Off-Peak Discounts
On-demand lease — the reservation the system creates when a GPU session is launched without a booking having been made. It draws on the Service Unit budget exactly as a booked window would; there is no free exploratory launch. → On-Demand Leases Charge Budget
OOMKilled — the pod status meaning the container reached its memory limit.
Frequently caused by the request/limit halving rather than by the number
requested. → Error Messages
Otter-Grader, Gradescope — grading tools used on the platform alongside nbgrader. → Choosing a Grading Tool
Overstay — a session continuing to run past the end of its guaranteed window. Nothing stops it; the time is no longer covered by what was booked, and the session has no protection. → What Ends a Session
Pinning — fixing a workspace to a dated image tag so that its members are not moved by a quarterly image update mid-term. → Pinning a Workspace
Pod — one running container as Kubernetes sees it. Pod IDs look like
ubellur-27068, and kubectl get pods lists the pods in an account's own
namespace.
Pod label (-l key=value) — an attribute attached to a pod at launch,
repeatable. This is how gpu-class is set.
Preemption — a session ending because the capacity it held was needed for somebody's booking. A capacity outcome, not a fault. → What Ends a Session
private/ — a per-user, cluster-wide directory. It appears unchanged in
every workspace and draws on the personal quota rather than any
workspace's. → Directories, Quotas & Cleaning Up
public/ — the workspace's shared area, readable by every member and written
by the grader account in a course. Draws on the workspace quota.
P3 / P4 — University of California data protection levels. P4 is prohibited here. P3 may be permitted after review, and vetting may take 4-6 weeks or longer. → Policy
Quota, GPU — the ceiling on how many GPUs of each class a workspace may hold at one time. Date-aware: it can be raised for a deadline week and revert on its own. Not the same as a Service Unit budget. → Quotas, Cohorts & Availability
Quota, storage — two separate pools. The workspace pool covers the
workspace home, public/ and teams/; the personal pool covers private/
and follows the member into every workspace.
→ Directories, Quotas & Cleaning Up
Research Cluster — the research-facing service, supported by Research IT at
rcd-support@ucsd.edu, with its own maintenance window: quarterly, at the break
between terms, about 7 days' notice, and all running jobs terminated.
Reservation — a booked window during which GPU capacity is held and a session is admitted ahead of the walk-up queue. A guarantee of access, not a running job — the session is still launched as usual. → Reservations
Reserve floor — capacity ITS keeps unborrowable in each GPU class. Idle is therefore not the same as available.
Resource tiers — the three limits that apply to a job: a single pod
(8 CPU / 32 GB / 1 GPU), the namespace in total (8 CPU / 64 GB / 1 GPU), and what
is available on request (up to 32 CPU / 128 GB). All three are defaults, not
maxima. → launch.sh Reference
rstudio-notebook — datascience-notebook plus the RStudio environment.
Derives from the CPU image and is not GPU-enabled.
Runtime limit — the deadline on a container: 6 hours by default, up to 12 if set at launch, longer only by arrangement. One of three clocks that can end a job. → The Runtime Limit
scipy-ml-notebook — datascience-notebook plus CUDA/GPU tooling,
TensorFlow and PyTorch. The default image for bare launch.sh.
Senior and junior borrowing — the seniority attached to borrowed capacity. Course workspaces always borrow senior; junior borrowing yields first.
Service Unit (SU) — a usage credit for GPU time. Each class carries an hourly rate, budgets are per-workspace and set administratively, and a reservation's cost is computed up front. → Service Units & Budgets
Slurm compatibility wrappers — sbatch, srun, squeue and scancel,
which translate into launch.sh. There is no Slurm scheduler behind them, and
no MPI or multi-node support. → Coming from HPC
Spawn — starting a browser session from a course's environment menu. A spawn failure is the message returned when it will not start, and it does not say why. → Sign-In & Session Problems
Standard image — one of the three images IT Services maintains:
datascience-notebook, scipy-ml-notebook and rstudio-notebook. They receive
priority support. → Standard Images
Team — a group within a workspace, holding data scoped to some of its members
— a project group, a lab bench, a capstone team. launch.sh -G list prints the
teams an account belongs to; -G <teamid> launches with that team's data
visible. Note -g is GPU count and -G is group.
→ Belonging to Several Workspaces
Team mode — a reservation feature in which teammates can act on one another's bookings. A teammate can cancel another member's booking, and that penalty cannot be waived. → Team Mode
teams/ — the directory where team-scoped data appears. Draws on the
workspace quota. → Directories, Quotas & Cleaning Up
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 course rosters auto-populate from. Auditors and observers are not on them and are added through Canvas. → When Access Starts & Ends
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 with
no root and no sudo, in a per-user Kubernetes namespace. sudo apt-get fails
by design. → The Hard Boundary
Workspace — the unit everything hangs off: one per course, lab, or catch-all
population. It anchors roster, storage, container images, GPU class access,
quotas and budgets. Named at launch with -W; listed by workspace --list.
→ What a Workspace Is
Workspace manager — the privilege tier held by an instructor, a TA or a PI. May view the group calendar, book on a member's behalf, and waive a cancellation charge; may not edit Service Unit budgets or group limits. → Managing a Group
If you still have questions or need additional assistance, email us at datahub@ucsd.edu or submit a ticket to the ITS Service Desk.