Research on PARASOL

This page is for researchers working without a lab workspace: graduate students, postdocs, undergraduate researchers, staff researchers, and faculty who have not yet established a group. It assumes familiarity with a shell.

Setting Up a Research Lab covers provisioning for a group. Projects & Independent Study covers for-credit coursework and capstones.

What the Platform Is Suited To¶

Datahub sessions and research work run on the same cluster, PARASOL (formerly DSMLP). Older pages and articles may call it the Research Cluster, and the work does not change with the name an article uses.

Suitable Work¶

The cluster suits interactive analysis, single-node GPU work at every GPU class size, fine-tuning, long-running batch jobs, and the broad range of research computing that does not require a supercomputer. It has operated since 2017 and carries both instruction and research.

Unsupported HPC Features¶

PARASOL is not a high-performance computing (HPC) system. It has no Slurm scheduler and no MPI or multi-node support. The sbatch-style wrappers are a compatibility layer rather than a scheduler. See Slurm Compatibility Wrappers. Coming from HPC covers HPC vocabulary and Slurm compatibility. Routing for work that requires these features is in Beyond PARASOL.

Obtaining Access¶

Request access through Research IT or rcd-support@ucsd.edu. Describe the work, its approximate resource requirements, and the timeframe.

The Working Environment¶

Signing In¶

Connect with ssh to the login node using Active Directory (AD) credentials. Duo applies. The VPN is not required for ssh, only for reaching a port inside a container. Connecting over SSH describes both.

The browser route at datahub.ucsd.edu is also available and is often the quickest way to inspect something. Access describes the routes onto the cluster.

The Login Node¶

The login node is for launching jobs and moving files, not for computation. See What the Login Node Is For. It is the one host every user passes through.

Launching a Job¶

A GPU job is launched with a wrapper script, for example:

launch-scipy-ml.sh -W <WORKSPACE> -c 8 -m 32 -g 1 -l gpu-class=large

workspace --list prints the ID to use for <WORKSPACE>; see Listing Workspace IDs.

Bare launch.sh and the wrapper scripts start with different default resources. See Default Resources. Limits apply per pod and per namespace, and larger allocations are available on request. See Resource Tiers and Administrative Requests. Resource Requests and Limits describes how the memory requested with -m relates to the amount guaranteed. launch.sh Reference describes the launcher.

Storage¶

Research home directories are substantially larger than course ones and are set up when access is provisioned. Workspace and Personal Quotas gives the sizes. Shared space is also available, as are optional mounts of external storage such as SDSC Universal Scale Storage. See Mounting External Storage. Workspaces & Storage describes the storage layout.

Software¶

The standard software images cover most requirements. Where they do not, a custom image may be built. Deriving it from a standard image, rather than starting from nothing, is the supported path. Environments covers images.

Root Access¶

Containers run unprivileged, with no root or sudo access, and dependency installation has to work within that constraint. See Root Access and System Packages.

Running Substantial Work¶

Job Modes¶

Job Modes describes the modes.

Runtime Limit¶

Jobs default to 6 hours, and up to 12 hours may be set at launch. A booking does not lift the limit. See Reservation Length and Session Runtime. The Runtime Limit describes the limit.

Checkpointing¶

Checkpoint long-running work. The runtime limit, idle culling, the end of a reservation window, and preemption can each stop a container independently of the code it is running. Checkpointing & Logging Long Runs describes the methods.

Monitoring and Session Events¶

Watching a Running Job covers observing a running job.

The reservation system reports on a GPU session with Kubernetes events, shown by kubectl describe pod, and by kubectl get events after the pod is gone. Reservation Events defines each:

• While a session waits: WaitingForReservation, ReservationFull, ReservationTooSmall, OnDemandLeaseDenied, OnDemandLeaseRejected, OnDemandAdmissionPaused, UnknownGpuClass, NoReservation, NoGpuRequest, AnnotationIgnored, NoMatchingNode, WaitingForNode.
• When it is admitted: RuntimeGuaranteed, ReservationRelinked, OverstayRelinked, BestEffortAdmitted.
• When it is stopped: Preempted, ReservationCancelled, ReservationReassigned.

Direct use of Kubernetes is available to advanced users. See Direct Kubernetes Use and Session Events.

Obtaining GPU Time¶

GPU Access describes the GPU allocation model in effect from Fall 2026.

The Shared Research GPU Pool¶

Researchers without a lab workspace draw on a school-wide shared research GPU pool. The pool provides a baseline allocation, with occasional boosts when capacity frees up, most notably over the summer, when instructional hardware would otherwise be idle. Service Unit (SU) budgets divide the shared pool equitably among the researchers drawing on it.

Choosing a GPU Class¶

Select the smallest GPU class the model fits within. Five classes are available. A larger class is not faster for a model that fits in a smaller one. It is scarcer and more expensive. GPU Classes describes the classes. Workloads by GPU Class gives examples of research work at each class, and the cards and slices behind it.

Multi-Day Reservations¶

A research workspace starts with the same length cap as a course workspace. Cluster administrators can raise it on the researcher's request, so that a multi-day fine-tuning run can be booked against guaranteed hardware. A member's own booking caps at 48 hours unless the workspace is in researcher mode. See Reservation Length Caps.

Borrowing¶

When a baseline allocation is exhausted, idle capacity may still be available through borrowing. See Borrowing Beyond Quota.

Launching Without a Reservation¶

On-Demand Lease Charges describes the charges. A launch holds a lease of 1 hour 10 minutes; for longer, use Extend in the reservation app at reserve.dsmlp.ucsd.edu.

Budget Windows¶

A research SU budget starts with the same size and weekly window as a course budget, and either can be changed on the researcher's request. See Budget Windows & Cadences. See Remaining Balance.

Cancellations and Missed Windows¶

Releasing a window at least 24 hours in advance costs nothing. A later cancellation, or handing back the end of a session, keeps the time used and usually part of the unused time. A window missed without cancelling is charged, and an individual researcher has no instructor from whom to request a waiver. The Cancellation Penalty describes the charge.

Off-Peak Hours¶

Off-peak discounts apply outside the peak evening hours. Work that can run at midday or overnight costs less and waits less. Peak & Off-Peak Hours describes peak and off-peak pricing.

Idle Culling¶

Idle culling reclaims a GPU session that stops using its GPU, on every GPU class and for research sessions as for any other. See What Counts as Idle.

Data¶

Moving Data¶

Browser upload suits small files. scp, sftp, rsync, git, and Globus handle substantial volumes. For the transfer routes, see Moving & Sharing Data.

Sharing Data¶

Inside the Workspace describes sharing data with collaborators, with a group, or publicly, and the permission model beneath it.

Restricted and Licensed Datasets¶

Some data available on the cluster carries usage terms. See Restricted & Licensed Datasets.

Data Classification¶

P4 data, such as clinical records and export-controlled information, must not be used on PARASOL. P3 data, which is legally or contractually protected information, may be permitted after review. Data Classification describes the classification levels and the review.

Publishing & Reproducibility¶

Reproducible Artifacts¶

An image tag, the code, and a checkpoint make a better reproducibility artifact than a home directory. See Building & Publishing a Custom Image.

End of Access¶

Retrieving Work Before Access Ends describes retrieval.

Support¶

Support for research users is routed by subject.

Urgent or broadly scoped problems may be escalated through the ITS Service Desk, with a plain statement of what is affected. See Response Targets. Getting Help describes support routing in full.

Beyond PARASOL¶

Work requiring tightly coupled multi-node computation, MPI, a batch scheduler, or capacity at a scale the cluster does not carry belongs on another platform. Research IT can connect researchers with campus and national resources. A group large enough to warrant its own guaranteed capacity is a matter for its PI. See Setting Up a Research Lab.

Report a problem with this page