.. _tutorials-onboarding-m1-quiz:

Quiz
====

Use this self-check to test your understanding of the CoSTAR cluster resources, usage etiquette, and baseline Linux environment.

These questions are browser-side self-checks. They are not formally graded and can be retaken as many times as needed.

.. raw:: html

   <br><br>

.. quizdown::

   ---
   primary_color: "var(--color-brand-primary)"
   secondary_color: "var(--color-background-secondary)"
   text_color: "var(--color-foreground-primary)"
   shuffle_questions: false
   shuffle_answers: true
   ---

   ### Based on the provided policies and resource descriptions, which of the following statements are TRUE?

   Select all that apply:

   - [x] It is forbidden to manipulate the scheduler to bypass queue limits or priorities.
        > Correct. Doing so will result in an immediate account suspension.
   - [x] The high-performance scratch area is temporary and is not backed up.
        > Correct. Both the Introduction and Usage Policy documents state that the scratch area is for active job data only.
   - [x] Connecting to the login servers using VS Code's remote connection features without caution can overload shared resources.
        > Correct. The policy warns users that unconfigured background server processes can rapidly degrade performance for everyone.
   - [ ] Your home directory has unlimited storage space for hosting massive deep-learning datasets.
        > Incorrect. Home directories have a strict 30GB quota. Production data must live in the WekaFS parallel filesystem.

   ### What action is recommended to prevent total data loss during long model training runs or simulations in the event of an unexpected disruption?

   1. [x] Implementing checkpointing to periodically save model states and data progress.
       > Correct. The policy advises using checkpointing to ensure you can resume runs smoothly from a recent state if a node experiences a disruption.
   1. [ ] Running the entire multi-day simulation directly on the login node for better stability.
       > Incorrect. Running execution scripts on the login node is a major policy violation.
   1. [ ] Requesting exactly double the memory allocation your code requires to create a virtual buffer.
       > Incorrect. Wasteful resource requests lower your Fairshare priority and keep you stuck in the queue.
   1. [ ] Storing all raw chunked training outputs in your default home directory.
       > Incorrect. Home directories have a small 30GB quota and will completely choke on large raw data streams.

   ### Why is it strictly forbidden to run computationally intensive experiments directly on the shared login node?

   1. [x] The login node is a shared system entry point with highly restricted resources; heavy tasks will crash or slow it down for all users.
       > Correct. The login node is exclusively for editing files, compiling small binaries, and managing your Slurm scripts.
   1. [ ] Because the compute nodes are physically slower and are only reserved for long-term file storage.
       > Incorrect. Compute nodes house your elite NVIDIA H200 GPUs and high-core processors.
   1. [ ] Because the login node lacks the encryption keys required to compile research code safely.
       > Incorrect. The restriction is entirely based on fair resource sharing and preventing system overloads.
   1. [ ] Because the Slurm scheduler automatically blocks the execution of binaries that do not contain a submission flag.
       > Incorrect. The scheduler cannot stop you from running a raw binary on the login node; it relies on monitoring policies and user compliance.

   ### After successfully logging into the cluster terminal for the first time, what will executing the command `echo $SHELL` output by default?

   1. [x] `/bin/bash`
       > Correct. Bash is the default environment shell provided across the cluster.
   1. [ ] `/bin/zsh`
   1. [ ] `/bin/tcsh`
   1. [ ] `/usr/bin/fish`

   ### Which of the following statements are TRUE regarding the high-performance capabilities of the CoSTAR cluster?

   Select all that apply:

   - [x] It is a fully parallel HPC supercomputer designed for distributed, multi-node scaling using MPI.
       > Correct! CoSTAR breaks legacy single-node restrictions.
   - [x] Nodes are interconnected via an ultra-low-latency 400 Gb/s InfiniBand fabric to support massive multi-node training routines.
       > Correct! This allows model training to scale seamlessly across multiple systems.
   - [ ] Workloads are restricted to running on a single isolated node and cannot scale out horizontally.
       > Incorrect. This was a limit on older facilities, but CoSTAR natively supports extensive cluster-wide parallel scaling.
   - [ ] Multi-GPU workflows can only leverage the 8 cards within a local chassis and cannot communicate over the cluster network.
       > Incorrect. The 400 Gb/s interconnect ensures that multi-node GPU communications run at peak performance.