Cluster Architecture

Overview

Questions

• What are partitions and why does Sagehen HPC have three?
• What hardware is in each partition?
• How do I view cluster status?

Objectives

• Understand the concept of partitions in HPC clusters
• Learn Sagehen’s specific hardware specifications for each partition
• Use sinfo to query cluster status

What are Partitions?

---

A partition in SLURM is a logical grouping of compute nodes with shared characteristics. Sagehen uses partitions to organize nodes by hardware type, time limits, and priority.

Think of partitions like different sections of a library: fiction, non-fiction, reference. Each section is organized for its purpose.

How Sagehen HPC’s partitions fit into the cluster: the head node dispatches jobs to the amd and gpu partitions, with the short partition for quick tests.

The “amd” Partition (Standard Compute)

The primary general-purpose partition for most research.

• 12 compute nodes
• Cores per node: 128 (AMD EPYC processors)
• Memory per node: 500 GB
• Interconnect: 100 Gb Infiniband
• Maximum job duration: 720 hours (30 days)

When to use: Most research computations, data processing, simulations, first-time job submissions.

The “gpu” Partition (GPU Acceleration)

For research requiring GPU acceleration.

• Multiple GPU-equipped nodes
• CPUs: 128 cores per node, 500 GB RAM
• GPUs:

GPU Type	Count	GPU Memory	Strength
NVIDIA A100	4	80 GB each	Excellent for training
NVIDIA L40S	4	48 GB each	Tensor-optimized / inference
NVIDIA RTX PRO 6000	2	96 GB each	Largest memory on the cluster (Blackwell)

That’s 10 GPUs in total across the gpu partition.

• Maximum job duration: 720 hours (30 days)

When to use: CUDA/cuDNN code, deep learning training, GPU-accelerated molecular dynamics.

Callout

GPU Allocation Matters

Request only the GPUs you need:

• Deep learning: Usually 1-2 GPUs per job
• Multi-GPU training: 2-4 GPUs (if your code supports it)
• Standard molecular dynamics: 1 GPU

Each GPU reserved limits availability for others. Be considerate!

The “short” Partition (Quick Jobs and Debugging)

For quick test jobs, debugging job scripts, and rapid prototyping.

• Shorter maximum walltime than amd and gpu (verify with sinfo -p short)
• Higher priority for short-duration work
• Useful for sanity-checking software and submission scripts before running long jobs

When to use: Debugging job scripts, quick test runs, single-task interactive debugging, and any work that finishes within the short partition’s walltime limit.

Viewing Cluster Status with sinfo

---

BASH

# See all partitions
sinfo

# More detailed output
sinfo -l

# Show only a specific partition
sinfo -p amd

# Show all nodes and their state
sinfo -N

Example output:

PARTITION  AVAIL  TIMELIMIT   NODES  STATE
amd           up 30-00:00:00     12  alloc
gpu           up 30-00:00:00      4  alloc
short         up    2:00:00      4  idle

(The exact short partition walltime limit may differ — always check sinfo -p short for current values.)

• AVAIL: Is the partition available (up/down)?
• TIMELIMIT: Maximum job duration
• STATE: Current state (alloc=some allocated, idle=free, down=unavailable)

Challenge

Challenge 1: Exploring Sagehen HPC’s Partitions

Query the cluster to understand its configuration.

Steps:

1. View partition information: sinfo -l
2. View node details: sinfo -N -l
3. Check the head node CPU: lscpu | head -20

Questions:

• How many nodes are in the “amd” partition?
• What is the maximum walltime for the amd partition?
• How many cores does a compute node have?

Show me the solution

• amd partition: 12 nodes
• amd partition maximum walltime: 720 hours (30 days)
• Each compute node: 128 cores

All standard compute nodes on Sagehen have identical configurations.

Real sinfo -l and sinfo -N -l output: the three partitions and per-node cores and memory.

lscpu on the head node — note it reports 64 threads on dual AMD EPYC 7313 chips, not the 128 cores of a compute node.

Key Points

• Partitions group nodes by hardware type and enforce time/resource limits
• Sagehen has three partitions: amd (general purpose), gpu (GPU-accelerated), and short (quick / test / debug jobs with a shorter walltime)
• Each compute node has 128 AMD EPYC cores and 500 GB RAM
• GPU partition offers 10 GPUs: A100, L40S, and RTX PRO 6000
• Use sinfo to check partition status, node states, and available resources
• Maximum job time: 720 hours (30 days) on amd and gpu; check sinfo -p short for the short partition limit