AICR Introduction

AICR Introduction

AICR is the AI Compute Resource: a GPU cluster that six Massachusetts universities built together and share, with 248 NVIDIA B200 GPUs and 152 RTX PRO 6000s. This lecture is the whole introduction in one film: what AICR is and how it relates to Northeastern's own Explorer cluster, how a project gets access, where your files live and for how long, how the partitions are cut, how you submit a job with Slurm, and what happens to your data when the project ends. Liam, in for Bear, narrates. Every number is checked against the AICR documentation and the live cluster.

Chapters 0:00 The idea: you run the big jobs on AICR, you develop elsewhere 0:16 Terms in this lecture 0:39 I. What AICR is 1:27 II. Getting on: the proposal, Open OnDemand, your username 2:01 III. Where files live: home, scratch, project 2:33 IV. Compute: partitions, batch vs devel, idle GPUs, sinfo 3:32 V. Running a job: sbatch, srun, Open OnDemand 4:45 VI. Globus, the yearly review, moving data home 5:21 Recap 5:50 Your turn

Your turn. Paste this into Claude: I'm about to run my first production job on a shared GPU cluster that uses Slurm. Here is my training setup: [describe it in two sentences]. Write an sbatch script that asks for one GPU and a 30-minute time limit. Then tell me which of my files belong in home, which in scratch, and which in the project directory. Then check two things yourself: does every #SBATCH line ask only for what your code can use? And is anything you can't afford to lose sitting in scratch?

Sources Northeastern Research Computing, AICR Introduction training (github.com/northeastern-rc-training/AICR-Introduction-2026); the AICR documentation; the AICR cluster itself, read-only, 2 October 2026. The sign-in page and the Globus documentation page shown are real captures from 2 October 2026. Questions about AICR: rchelp@northeastern.edu.

Made with the brutalist.art toolkit: Manim, Remotion and Kokoro, free and local. Narration is synthetic (Liam, in for Bear). This is an independent educational film, not an official Northeastern or AICR publication.

What this video covers

• 0:00 — The idea: you run the big jobs on AICR, you develop elsewhere
• 0:16 — Terms in this lecture
• 0:39 — I. What AICR is
• 1:27 — II. Getting on: the proposal, Open OnDemand, your username
• 2:01 — III. Where files live: home, scratch, project
• 2:33 — IV. Compute: partitions, batch vs devel, idle GPUs, sinfo
• 3:32 — V. Running a job: sbatch, srun, Open OnDemand
• 4:45 — VI. Globus, the yearly review, moving data home
• 5:21 — Recap
• 5:50 — Your turn

Transcript

A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.

The idea: you run the big jobs on AICR, you develop elsewhere

[0:00] Hello, this is Liam in for Bear. First, a correction. Aker is not the cluster where you build and test your AI work. It's the cluster where you run the big jobs. It's a shared GPU cluster for AI research in Massachusetts, and this lecture covers what it is, how you get on, where your files live, and how you run a job. Six terms: an HPC cluster,

Terms in this lecture

[0:18] Many computers wired together and shared by many people. The GPU, a chip that does AI math far faster than a normal processor. Slurm, the scheduler, you ask for resources and it decides when you run. A partition, a named group of machines with its own rules. A PI, the principal investigator, the faculty member who leads a project. An Open OnDemand, a website that gives you the cluster in a browser. Start with what it

I. What AICR is

[0:40] Is. Aker stands for AI compute resource. Six universities at the Massachusetts Green High Performance Computing Center, together with the Massachusetts AI Hub, built one cluster and they share it. Inside are two kinds of GPU: 248 Nvidia B200s and 152 RTX PR06000s. That's 400 GPUs in all. Around the GPUs, fast storage, a modern software stack for machine learning, and a CPU-only section for the work that needs no GPU, like cleaning and analyzing data. Northeastern already has its own cluster called Explorer. Treat Aker as an extension of it. You develop and test on Explorer. When the code works, you bring it to Aker for the production run. Aker is shared by every partner university, so it isn't the place to debug.

II. Getting on: the proposal, Open OnDemand, your username

[1:27] Getting on starts with a faculty member. The principal investigator submits a brief proposal through a request form. Access comes through that project. Once you're on a project, the easiest door is Open OnDemand at ood.icr.io. It sends you to a sign-in page. You choose your identity provider, then sign in with your Northeastern username and password. One thing changes, your username. On the cluster, it's slightly different, so accounts from the partner schools stay distinct. If you're j.smith on Explorer, you're j_smith_neu on Aether.

III. Where files live: home, scratch, project

[2:01] Your files live in three places, the same three as on Explorer. Home for small scripts in your environment, scratch for temporary output from running jobs, and the project directory for shared longer-term work. Home holds 100 GB. Delete a file there by accident, and you have 7 days to recover it. The project directory has the same 7-day safety net. Scratch is huge, 10 TB and up to 15 files, but nothing in scratch can be recovered, and there's a 30-day purge. So, move anything important out before it's gone.

IV. Compute: partitions, batch vs devel, idle GPUs, sinfo

[2:33] Compute is split into partitions. One is CPU only, four have GPUs, a batch and a dev partition for the B200s, and a batch and a dev partition for the RTX cards. Every GPU node carries about 2.3 TB of memory. Batch partitions are for jobs you submit and walk away from. They run up to 24 hours, and they support more than one GPU. Dev partitions are for interactive work, 4 hours at most, and four running jobs per user. Before you ask for more than one GPU, confirm your code can actually use them. If it can't, the extra cards sit idle. An idle is watched. A background process tracks GPU utilization and cancels jobs that sit idle. So, end your interactive session when you're done. A tool called job stats will also report how each job used what it asked for. You can see the live list yourself. The command sinfo prints every partition, its time limit, and its node count. Run it, and you may see more partitions than the five here, because the cluster keeps changing.

V. Running a job: sbatch, srun, Open OnDemand

[3:33] Aether runs Slurm, so there are three ways to work. sbatch for jobs that run without you, srun for an interactive shell, and open on demand in the browser. A batch job is a script. The lines that start with sbatch are your request, which partition, how many CPUs, one GPU, 16 GB of memory, and a time limit of 30 minutes. Below the request is the work clear the loaded modules, load CUDA and Mina Forge, activate your environment, then run your training script. Submit it with S batch and the file name. If your script writes logs into a folder, create that folder first. Then three commands follow the job

Around. SQ shows your queued and running jobs, S cancel stops one, and S act shows what a finished job actually used. For an interactive shell on a GPU node, use S run. This one asks the RTX devel partition for eight CPUs, 16 GB, and one GPU for 1 hour. Always set the time so the session ends even if you forget, and type exit to give the GPU back. Or skip the terminal. Open On Demand launches graphical sessions such as Jupyter Lab or a desktop. You choose the partition, the time, and the resources, and it submits the job to Slurm for you. Those sessions count against the devel limits.

VI. Globus, the yearly review, moving data home

[4:46] To move large data, the primary method is Globus. Northeastern has a subscription. You set up a Globus account with your Northeastern credentials, then connect a folder on Explorer or on your own machine to a folder on Aker. Last, the part people forget. Accounts and project directories on Aker are temporary. Once a year, there's a check-in. The PI renews the request and the project continues. And when the project ends, the data has to leave. Aker is not permanent storage. Move your results back to a Northeastern resource. Questions go to the research computing team at rc.help@northeastern.edu.

Recap

[5:21] Let's recap with Claude. Aker is a shared GPU cluster with 248 B200s and 152 RTX PR0 6000. You develop on Explorer and bring the production run to Aker. A PI's proposal gets you on and you sign in through Open On Demand. Home holds 100 GB. Scratch holds 10 TB with a 30-day purge. Batch partitions run 24 hours. Devel partitions run four. And idle GPU jobs get canceled. Projects are reviewed every year, and your data goes home at the end.

Your turn

[5:50] Your turn. Paste this into Claude. I'm about to run my first production job on a shared GPU cluster that uses Slurm. Here's my training setup. Describe it in two sentences. Write an Sbatch script that asks for one GPU and a 30-minute time limit. Then tell me which of my files belong in home, which in scratch, and which in the project directory. Then check two things yourself. Does every Sbatch line ask only for what your code can use? And is anything you can't afford to lose sitting in scratch? Aker introduction at Nick Bear Brown.