Joint HPC Exchange Constitutional Principles¶
The HPC service center is a steward of a commons.¶
It merely manage resources that are literally owned the stakeholders.
Policies must be consistent with Ostrom’s design principles¶
This ensures sustainable utilization of common pool resources.
Organizations have autonomy in structuring payment model.¶
Each organization can configure their computing budgets and resources, in a manner that is consistent with their research agendas and funding profiles, while being protected from exploitation by larger stakeholder organizations. Organizations, e.g. divisions, departments, institutes, collaborations and individual PIs, can provide their local community with: 1) free computing, 2) subsidized computing, 3) startup packages, or 4) “bridge funding” for PIs who are “between grants”.
Free computing is not provided by the service center.¶
Users must ask their funding agencies, collaborators, departments or deans to support their computing.
Policies for Stakeholders¶
All computing queues have associated rates and charges – there is no free computing for anyone, even stakeholders. Organizations can become stakeholders by purchase nodes. Stakeholders have priority access to their nodes via their dedicated queues. Stakeholder’s are required to make excess compute cycles available to all users via the shared queue. Under reasonable circumstances (e.g. grant or publication deadlines), stakeholders are allowed to temporarily block the shared queue from sending jobs to their hardware. Stakeholders are free (within broad limits) to decide on policies for their community within their queues. Lack of capacity on a dedicated queue is typically a signal to stakeholders that it’s time to invest.
Policies for Non-Stakeholders¶
Users who have no stakeholder affiliation are allowed to compute only on the shared queue. Users who have no stakeholder affiliation have no “right to compute”, but historically, this has never been an issue, as there has always been sufficient excess capacity available. Lack of capacity on the shared queue is a signal to non-stakeholders that it’s time to become a stakeholder – or compute elsewhere. Practically speaking, there is usually plenty of capacity on the shared queue.
The Shared Partition¶
Shared queue charges defray the operating charges of the stakeholders. This incentivizes sharing, e.g. if a stakeholder is not using their hardware, the shared queue charges (incurred by other users) will cover the operating expenses of the stakeholder’s node. Stakeholders use the shared queue for surge capacity when their own hardware does not suffice. Lack of capacity on the shared queue is typically a signal to stakeholders and non-stakeholders alike that it’s time to invest in nodes. Historically, the cost for using the shared partition is about $0.01 per hour for a job using a single core and 5GB of RAM.
Salient storage policies¶
There is no free storage Each user account comes with 100GB of home directory space (current rate is $0.45/GB/year, so $45/yr MAX. Allocated (guaranteed) storage rate is between $18-$25 TB/year limited ‘leased’ storage is available. Recent rate is ~$44/TB/year (based on TB actually used) Home directories are backed up by default, all other storage is backed up upon request.