Policy

Check the conditions for using Datahub and PARASOL (formerly DSMLP) before you put data on the platform, host a service that people outside it can reach, or set a deadline that could fall in a maintenance window.

Data Classification¶

P3 and P4 are University classification levels. The University's classification is the authority on which level applies to a dataset.

P4 Data¶

P4 data is prohibited. Highly sensitive information, such as clinical records or export-controlled material, must not be placed on Datahub or PARASOL. There is no review path and no exception.

P3 Data¶

P3 data may be permitted after review. Legally or contractually protected information is not categorically excluded, but it must be raised with the contacts in Contacts for Classification Questions before any of it reaches the cluster. Depending on the nature of the data, vetting may take 4-6 weeks or longer.

Examples by Kind of Data¶

ITS documentation gives these examples:

Contacts for Classification Questions¶

For instruction and coursework, email parasol@ucsd.edu. For research, email Research IT at rcd-support@ucsd.edu. See Restricted & Licensed Datasets. Getting Help lists other support routes.

Appropriate Use¶

IT Acceptable Use Policy¶

The campus-wide IT Acceptable Use Policy applies to Datahub and PARASOL as it does to any campus IT resource, including its prohibitions on commercial or political activity, hacking or cyberstalking, and other types of unwelcome behavior. The platform's academic purpose does not relax it.

Individual Access and Sharing¶

Access is granted to an individual, not to that individual's collaborators. An account, its home directory, and any dataset granted with it are for the account holder's own use and are not passed on. The workspace's public/ and teams/ areas are the sanctioned route for sharing. See Inside the Workspace.

Hosting Externally Available Services¶

Datahub and PARASOL should not be used to host externally available services or applications, except as required for coursework or projects. A class demo, a project web application, or a service an assignment requires is in scope. A production service, a persistent public endpoint, or anything whose outage someone outside the University would notice is out of scope. Reaching a Notebook or a Service describes access to a service running in a container.

Availability and Reliability¶

A compute node can be drained for patching, a pod can be preempted, and a session has a deadline. These conditions apply primarily to the compute nodes that execute user jobs. Critical components such as networking, file storage, and backups are maintained to Enterprise IT standards.

Use of Shared Resources¶

The practices in this section are requests rather than rules, with the exception of the prohibition in Computation on the Login Node.

Stopping Unused Sessions¶

Sessions that are not in use are expected to be stopped. GPU cards are assigned to containers exclusively, and a card attached to an idle session is unusable by any other user. Logging out, closing the tab, or closing a laptop leaves a session running; stop it with File → Hub Control Panel → Stop My Server, or delete the pod from the login node. See Stopping a Session.

Idle culling is not a substitute for stopping a session. On a quiet cluster, the delay before it acts is measured in hours. What Counts as Idle describes the idle criteria.

Debugging on CPU¶

Code is expected to be debugged on CPU before a GPU is requested. PyTorch and TensorFlow both switch easily between CPU and GPU. Access describes the ways to start a session.

Sizing Resource Requests¶

Requests are expected to name the smallest resources that work: the smallest GPU class a model fits in, the memory a job needs rather than the maximum permitted, and the hours the job actually takes. A larger GPU class is not faster for a model that already fits in a smaller one. GPU Classes describes the classes.

Computation on the Login Node¶

Running computation on the login node, such as a training script, a build, or a long analysis, is prohibited; the login node is for launching jobs and moving files. See What the Login Node Is For.

Personal Copies of Shared Datasets¶

A shared dataset is expected to be read where it lives rather than copied. A per-student copy of a large corpus exhausts a home quota. Datasets describes shared datasets.

Canceling Unused Bookings¶

A booking that will not be used is expected to be canceled. Canceling at least 24 hours ahead costs nothing, and the capacity returns to the pool. The Cancellation Penalty describes penalties.

Peak GPU Demand¶

Demand for GPUs may exceed capacity at peak hours during 10th and Finals Weeks, or at assignment deadlines. Work shifted to a daytime or off-peak hour costs less. See Peak & Off-Peak Hours.

Self-Supporting Programs¶

Self-supporting programs such as MAS and MBA may use Datahub and PARASOL for coursework or projects. UC policy requires ITS to recover the associated direct and indirect costs. To discuss cost recovery, email parasol@ucsd.edu.

Workspace and Personal Quotas sets out charges for storage.

See also: Mounting External Storage

Scheduled Maintenance¶

Weekly Maintenance¶

Weekly maintenance runs Tuesdays, 6-8 AM, across the whole cluster, for courses and research alike. The work is usually limited to a subset of worker nodes, in which case running jobs are unaffected. Do not plan a deadline within this window. Infrequent 'Critical' updates may require downtime outside the window; ITS notifies instructors of such downtime as soon as is practical.

Quarterly Maintenance¶

The quarterly maintenance window falls at the break between terms, with about 7 days' notice. All running jobs are terminated, for courses and research alike. Maintenance Closures describes the effect on sessions.

Report a problem with this page