From HPC Fundamentals to Research
Workflows with AI Agents
Dr Zhipeng (Zippo) He
School of Mechanical, Medical and
Process Engineering
Queensland University of Technology
Use the demonstrations and discussion to relate the workflow to your own research.
Follow the linked tutorials afterwards for step-by-step instructions.
What are you trying to find out,
and what does your computer need to do?
Keep that task in mind as we meet five researchers.
“And I still have to
write the revision…”
“Add baselines, ablations and repeated runs.”
The experiments run one at a time.
Interpretation and writing still need time.
“It fits on disk.
Why won’t it run?”
The program needs working memory.
Reading a file is only the beginning.
“I can test the code.
But not the real setup.”
What changes when components run on different machines?
Local processes can test the code.
Messages cross a network.
Each machine has its own CPU and RAM.
“My workday cannot
wait for this to finish.”
“I changed my copy.
Did you get the update?”
PS C:\Users\user> ssh aqua[user@aquarius01 ~]$ qsub -I \ -P YOUR_RPID -m abe \ -l select=1:ncpus=1:mem=1GB \ -l walltime=00:05:00qsub: waiting for job 12345678.aqua-pbs to startqsub: job 12345678.aqua-pbs ready[user@cpu1n001 ~]$ hostnamecpu1n001[user@cpu1n001 ~]$ echo "hello from my job"hello from my job[user@cpu1n001 ~]$ exitqsub: job 12345678.aqua-pbs completed[user@aquarius01 ~]$
“Why are values missing in my data?”
Inspect the data, try a command,
then decide what to do next.
“Let preprocessing run while I write.”
Lily submits her prepared script.
It runs and saves results without further input.
You inspect the data and choose the next step.
Lily can keep writing while the script runs.
Both interactive and batch jobs can use CPUs or GPUs.
A PBS file is a plain-text shell script containing a resource request and the commands to run.
#!/bin/bash
#PBS -N hello
#PBS -P YOUR_RPID
#PBS -l select=1:ncpus=1:mem=1GB
#PBS -l walltime=00:05:00
#PBS -j oe
#PBS -m abe
cd "$PBS_O_WORKDIR"
hostname
echo "hello from my job"qsub hello.pbsLines beginning with #PBS
On the compute node, once resources are allocated
cd selects the submission directory.
hello.o<job-number>cpu1n023 hello from my job
Learn from reviews.
Generalise to new ones.
Fine-tune DistilBERT, a pretrained text model, using reviews labelled positive or negative.
Two hours of my life I will never get back. The plot made no sense and the acting was wooden.
Try one epoch, estimate the time,
then run the full task in batch.
A Python interpreter and the installed libraries that the program imports.
Plus the libraries for data and evaluation
The trial and batch job use this environment.
Record the libraries, including GPU-enabled PyTorch,
in the project’s dependency files.
pyproject.tomlRequired packagesuv.lockResolved versionsuv syncInstall the environmentFetch the pretrained model and reviews without training.
uv run python imdb_sentiment.py --fetch-onlyuv run selects this environment; training reads inputs offline.
qsub -I -P YOUR_RPID -m abe \
-l select=1:ncpus=6:ngpus=1:mem=32GB:gpu_id=H100 \
-l walltime=02:00:00One H100 MIG slice (~10 GB GPU memory). Exit when finished.
mkdir -p runs/trial
uv run python imdb_sentiment.py \
--epochs 1 --batch 16 --seed 42 --lr 2e-5 \
--model-dir runs/trial/model \
--out runs/trial/results.jsonSame full dataset and settings as the batch job.
Only the epoch count changes; trial outputs stay separate.
runs/trial/model/ + results.jsonCheck CUDA, the completed epoch and saved files.
Record elapsed time and the GPU slice used.
workshop.pbs#!/bin/bash
#PBS -N imdb_demo
#PBS -P YOUR_RPID
#PBS -l select=1:ncpus=16:ngpus=1:mem=64GB:gpu_id=H100
#PBS -l walltime=01:00:00
#PBS -j oe
#PBS -m abe
cd "$PBS_O_WORKDIR"
mkdir -p "runs/demo"
uv run python imdb_sentiment.py \
--epochs 3 --batch 16 --seed 42 --lr 2e-5 \
--model-dir "runs/demo/model" \
--out "runs/demo/results.json"16 CPUs · 64 GB RAM · 1 H100
1 hour as a conservative starting limit.
Refine after the full-H100 run.
cd enters the submission directory.uv run uses the project’s .venv.
No separate activation is needed.
Save the model and results inruns/demo/.
Q means queued. PBS accepted the job, but execution has not started.
16:14 is the current estimate. It can change as the schedule changes.
The scheduler’s comment reports insufficient available GPUs for this request.
runs/demo/results.jsonCounts, settings, accuracy
and resource observations
runs/demo/model/Saved model and tokenizer
for classifying new reviews
imdb_demo.o<job-number>Progress, errors and PBS summary
less imdb_demo.o12345678 cat runs/demo/results.json
#PBS -m abe... {'epoch': 3, ...}
accuracy 0.9161 on 25,000 test reviews
done ... in 480.0s -> runs/demo/results.jsonPBS Job 12345678.aqua Wall time : 00:08:37 Mem usage : 1272032kb
{
"job_id": "12345678.aqua",
"train_reviews": 25000, "test_reviews": 25000,
"epochs": 3, "device": "cuda",
"gpu": "NVIDIA H100 80GB HBM3",
"accuracy": 0.9161
}Use the run’s measurements to choose what to keep, reduce or increase.
Too little memory or walltime can stop the job.
Leave room above the observed need.
Allocated CPUs, RAM and GPUs are unavailable to other jobs while yours runs.
Keep the headroom that helps the GPU work.
PBS must find the requested hardware for the requested duration.
Extra size or time can rule out an opening.
Requested size, fair share and eligible waiting affect the score. Smaller does not automatically rank higher.
Hardware and walltime must fit the available opening. A shorter request may fit before reserved work.
For the task you had in mind:
What does your research require that your current setup cannot support?
Job arrays, a larger-memory allocation, communicating processes across nodes,
unattended batch execution or shared storage.
Identify one step you would need help understanding or carrying out.
Reminder: Choose HPC for the gaps it can address in your research,
not simply for the power it offers.
Inspect accessible scripts and logs directly.
Compare the evidence, make an agreed edit, check it.
Prepare scripts, configurations and outputs within scope.
Learn why a job stopped and what to check next.
Less operational work to carry. More room to understand and direct it.
Agree the work. Give room to act. Review the evidence.
An analogy for direction and review. People and agents differ in understanding and accountability.
GPT
Claude
Gemini
DeepSeek
Qwen
Llama
Mistral
Grok
Gemma
Kimi
GLM
Command
Jamba
Nova
MiniMax
Hunyuan
ERNIE
Yi
InternLM
Baichuan
OLMo
Falcon
Nemotron
Granite
Solar
Liquid
Reka
Cogito
MiMo
LongCat
StepChatGPTClaude appGemini app
Perplexity
Poe
NotebookLM
Meta AI
Doubao
YuanbaoKimi app
Kagi Assistant
Pi
LM Studio
Ollama
Open WebUI
Cherry Studio
LobeHub
Monica
YouMind
Codex
Claude Code
Cursor
Gemini CLI
GitHub Copilot
Cline
Devin
Kilo
Kiro
Junie
OpenCode
goose
OpenHands
Antigravity
Replit Agent
Lovable
v0
TRAE
Qoder
Zencoder
Manus
Skywork
Flowith
Dify
Coze
n8n
Zapier Agents
CrewAI
Hermes Agent
OpenClawThe products change. The working principles transfer.
Learn to integrate an agent into your research pipeline and HPC: connect it, direct its work, and check the results.
Install your chosen application on your computer.
Open the local folder for your research project.
Keep the agent application off Aqua’s login nodes.
Sign in, or configure a supported provider and API key
in the application’s settings. Choose a capable model.
An affordable coding model is a useful starting point.
Keep the private key local; register the public key on Aqua.
Test SSH yourself, then allow the agent to use it.
Start with manual approval of remote commands.
ssh <user>@aqua.qut.edu.auAsk the agent to read Walltime Chronicles’ agent index.
Attach relevant Aqua docs and provide project paths,
the environment and a working PBS script.
State permitted actions and where outputs should go.
WebFetch(https://zhipenghe.me/Walltime-Chronicles/llms.txt)# Walltime Chronicles Content Summary ## Full Index of Links and Descriptions …
WebFetch(https://zhipenghe.me/Walltime-Chronicles/agents/aqua-guide.txt)Read(<project>)EISDIR: illegal operation on a directory, read '<project>'
I appreciate your request, but I need to clarify what you’re asking me to do. … returning the entire document word-for-word would violate the constraint you’ve also specified …
WebFetch(https://zhipenghe.me/Walltime-Chronicles/agents/aqua-guide.txt)WebFetch(https://zhipenghe.me/Walltime-Chronicles/cheatsheet/)# Aqua HPC Cheatsheet - Structured Reference ## PBS/QSUB Commands …
# Structured Extraction: QUT Aqua Guidance ## (1) System Overview Facts …
Write(<project>\docs\aqua-guide.md)File created successfully at: <project>\docs\aqua-guide.md
Read the index, then the Aqua guide.
After a failed read and limited web response, fetch a summary and cheatsheet.
Saved docs/aqua-guide.md with rules and source links. Review before reuse.
Same code. Same Aqua workflow. Agent assistance at each step.
Read the project.
Approve the action.
Check the evidence.
Less manual handoff. You stay in control.
Bash(ssh -o BatchMode=yes -o ConnectTimeout=15 aqua "cd ~/hello-aqua && sed -n '1,230p' pyproject.toml")[project] name = "hello-aqua"
Bash(ssh -o BatchMode=yes -o ConnectTimeout=15 aqua "cd ~/hello-aqua && sed -n '1,230p' imdb_sentiment.py")MODEL = "distilbert/distilbert-base-uncased" DATASET = "stanfordnlp/imdb"
Bash(ssh -o BatchMode=yes -o ConnectTimeout=15 aqua "cd ~/hello-aqua && sed -n '1,230p' workshop.pbs")uv run python imdb_sentiment.py \ --epochs 3 --batch 16 --seed 42 --lr 2e-5 \ --model-dir "runs/demo/model" \ --out "runs/demo/results.json"
Name the project and run. Ask before submission.
SSH returns file contents that may enter model context.
The code defaults to two epochs; this job requests three.
Check the environment, settings, model path and result path.
Bash(ssh aqua "cd ~/hello-aqua && cat workshop.pbs")#PBS -l select=1:ncpus=16:ngpus=1:mem=64GB:gpu_id=H100 #PBS -l walltime=01:00:00 #PBS -j oe #PBS -m abe cd "$PBS_O_WORKDIR"
Bash(ssh -o BatchMode=yes -o ConnectTimeout=15 aqua "cd ~/hello-aqua && qsub workshop.pbs")12345678.aqua
Bash(ssh -o BatchMode=yes -o ConnectTimeout=15 aqua "qstat 12345678")12345678.aqua imdb_demo Q (queued)
Bash(ssh aqua "qstat 12345678")12345678.aqua imdb_demo R (running)
ssh -o BatchMode=yes -o ConnectTimeout=15 aqua "cd ~/hello-aqua && qsub workshop.pbs"
Submit the reviewed IMDb job on Aqua
Do you want to proceed?
Check settings, resources and output paths.
Review the full SSH command before submission.
Submission joins the queue; it does not start compute.
It saves outputs on Aqua, even after the conversation ends.
Bash(ssh -o BatchMode=yes -o ConnectTimeout=15 aqua "cd ~/hello-aqua && cat runs/demo/results.json")"train_reviews": 25000, "test_reviews": 25000, "epochs": 3, "device": "cuda", "gpu": "NVIDIA H100 80GB HBM3", "accuracy": 0.9161, "elapsed_s": 480.0
Bash(ssh -o BatchMode=yes -o ConnectTimeout=15 aqua "cd ~/hello-aqua && tail -n 90 imdb_demo.o12345678")accuracy 0.9161 on 25,000 test reviews done accuracy 0.9161 in 480.0s -> runs/demo/results.json, model in runs/demo/model/ Wall time : 00:08:37 Exit_status : 0
Read the result and matching log, so evidence from different runs is not mixed.
Evaluated on 25,000 test reviews after three epochs.
Exit status 0. Program timing: 8 min.
Report the final saved evaluation: 91.61%.
Explain what is known and what still needs checking.
State the outcome, what must stay the same
and where the agent should stop.
“Keep the experiment settings.
Prepare the job. Ask before submission.”
Configure supported tool permissions.
Begin with approval for remote changes and jobs.
Read the project → permitted
Submit a job → approve the exact action
Review consequential changes and returned evidence.
Interrupt repeated edits without a diagnosis.
“Stop editing. Show the error, your changes
and what is still uncertain.”
You decide what is worth doing
and what the results mean.
Remember to include in your paper:
“We acknowledge support from the Queensland University of Technology (QUT) through its provision of subsidised access to specialised research infrastructure including HPC and the eResearch team at QUT for providing expertise in enablement of this project.”
OpenAI GPT-6 AstraClaude Fable 5.1 & Opus 5Workshop materials developed with assistance from these models. Content reviewed and final decisions made by the presenter.
Processors do the work. Memory is the workspace; storage keeps the files.
CPU analysis
and simulations
Request only the CPU cores your program can use.
Aqua CPU node
192 physical cores · 1.5 TB RAM
Software that
uses a GPU
Check GPU support and the GPU memory your workload needs.
Aqua GPU memory per full card
H100: 80 GB · A100: 40 GB
More RAM on
one machine
Use this when your workload needs more RAM than a standard node provides.
Aqua large-memory node
6 TB RAM · 192 physical cores
Installed capacities shown. Usable resources are lower; interactive limits differ. GPU memory is separate from system RAM.
An RPID identifies the research project in QUT’s Data Management Planner (DMP).
Sign in at data-mgmt-plan.qut.edu.au.
Existing project: ask the plan lead to add you as a collaborator.
New project: Start Data Management Plan → fill in details → Save.
Use a project ID you have access to.
Replace YOUR_RPID in the submission.
Project ID ≠ job ID. DMP identifies the project. PBS assigns a new job ID for each submission.
In a batch job’s PBS header.
Or supply it when submitting:
qsub -P YOUR_RPID job.pbs
Uppercase -P · use the RPID for this research.
qsub -I -P YOUR_RPID \
-l select=1:ncpus=1:mem=1GB \
-l walltime=00:05:00-I means interactive.
One CPU core, 1 GB RAM,
up to five minutes after the job starts.
hostname
echo "hello from my job"
exithostname shows where commands run.echo prints a message in the session.exit releases this allocation.
← Back to the allocation demonstration
Already logged in to Aqua. Replace YOUR_RPID with your project ID. This request is for the tiny example shown.
Each process produces its own research result.
No communication between runs is required.
The software launches and coordinates these components.
A cluster network may differ from the intended deployment.sweep.pbs · array-specific lines#PBS -J 1-10%3
#PBS -m a
cd "$PBS_O_WORKDIR"
read -r lr seed < <(
sed -n "${PBS_ARRAY_INDEX}p" configs.tsv
)
run="runs/sweep/$PBS_ARRAY_INDEX"
mkdir -p "$run"
uv run python imdb_sentiment.py \
--epochs 3 --batch 16 --lr "$lr" --seed "$seed" \
--model-dir "$run/model" \
--out "$run/results.json"Keep the resource directives from the revised job. Each member receives that allocation.
configs.tsv has ten rows, no header:
learning rate, then seed.
The index names the output directory.
Abort-only mail avoids a message for every start and finish.
configs.tsvLearning rate and seed for each rowPBS member logsExit status and whole-job runtimeruns/sweep/<index>/results.json and the saved modelAccount for every planned member.
Missing or failed runs stay visible.
Inspect failed logs before retrying.
Compare completed runs under the same protocol.
Sam’s campaign: a checked program, a justified request, and identifiable results for every configuration.
The same data and settings, from one epoch to three.
#PBS -l walltime=HH:MM:SSQueue waiting is separate. Keep an allowance; do not subtract an assumed full-GPU speedup.