Cloud HPC and AWS ParallelCluster

This file focuses on cloud HPC design, especially on AWS.

Why Cloud HPC?

Cloud HPC is useful when you need:

• burst capacity
• rapid provisioning
• flexible instance choices
• project-based clusters

Cloud HPC is less attractive when you need:

• stable ultra-high utilization
• maximum low-latency performance at all times
• strict on-prem data locality

Cloud HPC Design Questions

Before choosing tooling, ask:

• Is the workload tightly coupled?
• Is it independent batch?
• Does it checkpoint?
• Is spot acceptable?
• How large is the dataset?
• How expensive is data movement?

AWS ParallelCluster

AWS ParallelCluster is a cluster deployment and management tool for HPC on AWS.

It commonly manages:

• head node
• Slurm scheduler
• compute fleets
• storage integration
• EFA-enabled networking

Typical ParallelCluster Architecture

Users
  |
  v
[Login / Head Node]
  |- Slurm
  |- Cluster config
  |
  v
[Compute Queues]
  |- CPU
  |- GPU
  |- Spot
  |- On-demand
  |
  v
[Storage]
  |- FSx for Lustre
  |- EBS
  |- EFS
  |- S3

Storage Mapping on AWS

Compute and Network Mapping

Tightly coupled MPI

Prefer:

• homogeneous instance types
• EFA-enabled instances
• placement groups
• on-demand first unless restart behavior is strong

Independent batch

Prefer:

• flexible instance pools
• spot-heavy strategies
• simpler storage path

GPU training

Prefer:

• homogeneous GPU generations per queue
• locality-aware launches
• strong checkpointing plan
• fast dataset access

Spot vs On-Demand

Use spot when:

• jobs are restartable
• checkpointing exists
• tasks are independent or interruption-tolerant

Avoid spot when:

• jobs are tightly coupled
• restart is expensive
• deadlines are strict

Cost Governance

Cloud HPC often fails from cost drift, not technical impossibility.

Add:

• project tagging
• budget alarms
• queue-specific policies
• idle resource cleanup
• explicit spot/on-demand rules

Interview Summary

Strong cloud HPC answers show:

• correct workload classification
• service choice justified by workload, not by brand familiarity
• storage and networking mapped explicitly
• cost and interruption strategy included

Related Files

• HPC.md
• HPC-01-Fundamentals.md
• HPC-03-Storage-Networking-Operations.md
• HPC-05-Interviews.md