College of Information ScienceComputing Resources Working Group · Point people: Barney Maccabe and Greg Chism
Run benchmarks¶
Supplemental reference · Browse supplements · Current report
Supporting compute pilot. These DS/MLIS proposals retain their own $0 cash and Greg’s four-hour weekly envelope. They are not college-wide capacity findings or the working group’s selected solution. Return to working-group deliverables →
Measure before piloting. Runnable DS and MLIS components exist; provider runs remain pending. Local synthetic controls are not AWS or CUDA evidence.
| Component | Run on | Record |
|---|---|---|
| CPU data-science and MLIS tests | Already-authorized campus UITS/AWS, component testing only | Wall time, configured threads, memory, input/output sizes and failures |
| GPU DS and MLIS notebooks | Kaggle or free Colab before piloting | Actual device/VRAM, runtime, allocated memory, quota/session limits, failures and exports |
| Target replay | Awarded ACCESS CPU/GPU resource | Equivalent outputs, actual architecture, provider accounting and concurrent behavior |
Download runnable benchmark bundle · Benchmark result template
Use the results¶
- Agree on inputs, configuration and correctness checks with each course lead.
- Run warm-up and retained repetitions. Preserve failed runs and platform labels.
- Export results, environment and hardware details.
- Use the included
size_access.pyand plan template for component demand. Add interactive idle time, provisioning, staff, storage, retries and concurrency without double counting. - Confirm target-provider billing units. CPU core-hours and GPU device-hours are not automatically ACCESS credits.
The MLIS GPU component is a projection/retrieval proxy; it does not establish pretrained semantic-retrieval or OCR quality. Validate those tasks separately if required.