You started a job on your laptop on Monday. It's Thursday and the progress bar hasn't moved much. Maybe it's a simulation you need to run a few thousand times, a model you want to score across every file in a dataset, or a genomics pipeline that eats memory for breakfast. Your campus doesn't have a research computing office you can walk into, so the usual advice is to wait, or to cut the analysis down until it fits the machine you own.
Here's what most researchers never get told: the United States runs national computing systems that are free to use if you're at a US academic or nonprofit institution, and you don't need your own grant to qualify. A graduate student can apply. The catch isn't money or eligibility. It's that two of the best options, the OSPool and NSF ACCESS, solve different problems, and picking the wrong one is where most first attempts stall. This is a plain guide to telling them apart and getting started.
One question sorts most of it out
Before you look at any signup page, describe your work in one sentence: is it many independent runs, or one big run?
Many independent runs means the same program executed over and over with different inputs, and no single run needs to talk to the others while it works. A parameter sweep, a bootstrap with ten thousand resamples, scoring a model across a folder of images, running one analysis per subject or per gene. This pattern is called high-throughput computing, and it's what the OSPool is built for.
One big run means a single calculation that has to stay together: a large simulation whose pieces exchange data every step, a model that needs a big GPU or far more memory than a laptop has, anything that won't split into independent chunks. That's the traditional supercomputer job, and it points you toward ACCESS.
Plenty of projects have some of both, and that's fine. Start with whichever describes the part that's stuck right now.
The OSPool: high throughput, no proposal to write
The OSPool is a national pool of computing capacity run by the OSG Consortium and funded by the National Science Foundation through the PATh project. Campuses across the country contribute the machines; researchers share the pool. If your work is affiliated with a US academic, government, or nonprofit organization, you can use it at no cost.
What makes it worth knowing about: there's no allocation proposal. You don't write a case for your project or wait on a review committee. You request an account on an OSG-operated access point, and once your workflow is packaged as many small jobs, the pool runs them wherever capacity opens up. For work that splits cleanly into independent tasks, this is usually the fastest door in, and the ceiling is high. People routinely run tens of thousands of jobs.
The tradeoff is that the OSPool rewards a particular shape of work. If your job can't be broken into pieces, or each piece needs a specialized GPU and constant communication with the others, this isn't the right tool. Get an account through the OSG Portal or an access point, and if your analysis is already a loop over files or parameters, you're most of the way there.
NSF ACCESS: allocations on the national supercomputers
ACCESS (the letters stand for Advanced Cyberinfrastructure Coordination Ecosystem: Services and Support) is the NSF program that hands out time on national supercomputers and data systems. It replaced the older XSEDE program. The word "allocation" makes people picture a grant competition, and for the entry level that picture is wrong.
ACCESS has four request sizes, and the smallest two are quick:
- Explore. A one-page abstract and your CV. Approved in days, not months. It's meant for benchmarking your code, testing that it runs, and pilot experiments. This is where almost everyone should start.
- Discover. A one-page proposal, good for a modest, ongoing research activity once you know your code works.
- Accelerate. Up to a three-page proposal that lays out how you'll use the resources. This is for an established project with real scale.
- Maximize. The largest awards, reviewed on a set schedule a couple of times a year. You apply here when you already know exactly what you need and why.
Awards for the first three tiers come as ACCESS Credits, a kind of currency you spend on the specific systems you choose from the menu. That menu includes ordinary clusters, big-memory machines, and GPU systems that suit machine learning and AI work. You apply at allocations.access-ci.org, and Explore requests can go in any time. To qualify you need to be a researcher or educator at a US academic or nonprofit institution, and again, you don't need to hold an NSF grant.
Explore takes a one-page abstract and a CV and clears in days. If you've been telling yourself national supercomputing is out of reach, that's the whole barrier.
A short decision guide
If your work is a loop over many independent inputs and you want to start this week without writing anything, go to the OSPool. If you need a big GPU, a lot of memory, or a single tightly-coupled simulation, get an ACCESS Explore allocation and pick a resource that matches. If you're doing machine learning and want GPU time, ACCESS is usually the better fit, and Explore is enough to benchmark before you ask for more. And if you genuinely can't tell, apply for an Explore allocation anyway. It costs you an afternoon and gives you a place to test.
What to have ready before you apply
Neither path asks for much, but a little preparation saves a round trip. Have a couple of sentences describing your project in plain language, an up-to-date CV or short bio, and a rough sense of what you'll actually run: roughly how many jobs or how many hours, whether you need a GPU, and how much data moves in and out. You don't need precise numbers. A reviewer or a facilitator would rather see an honest estimate than a blank.
The mistakes that cost people a semester
- Choosing the venue before looking at the workload. Forcing a tightly-coupled simulation onto the OSPool, or splitting a naturally parallel job onto a single big node, wastes the resource and your time. The one-sentence test above prevents most of this.
- Waiting for a "real" allocation. Explore exists so you can start small. Benchmark first, then ask for size once you know your numbers.
- Skipping the package step. Getting on the system is the easy part. Turning your R or Python analysis into something a scheduler can run, a container or environment file, a submit script, a plan for your data, is the part that actually needs attention. Budget for it.
- Going it alone when you don't have to. Both ecosystems have people whose job is to help, and using them is not cheating.
Where to get help
You don't have to learn this from documentation alone. The Campus Champions network connects staff at hundreds of institutions who volunteer to walk colleagues onto national resources. The OSG Consortium runs a free annual school that trains researchers and facilitators hands-on. ACCESS has its own support and facilitation program for people who are new to allocations. None of these ask about the size of your institution before they help.
The takeaway
The national computing that would finish your stuck analysis overnight already exists, it's paid for with public money, and it's open to you. The only real decision is which system fits the work: the OSPool for many independent jobs with no proposal, ACCESS for the big single jobs and GPU time, starting with a one-page Explore request. Pick one, get an account this week, and stop shrinking your research to fit a laptop.
Getting a first researcher running on these systems is a lot of what I do. If you or your team want help choosing the right venue and getting a workflow onto it, get in touch and we can scope a short pilot.