Module 4: Specialized Infrastructure

⚡ High Performance Computing (HPC)

Component	Purpose	Details
Cluster Placement Groups	Low-latency networking	Instances in same rack, 10–25 Gbps enhanced networking
Elastic Fabric Adapter (EFA)	OS-bypass networking	MPI, NCCL for tightly coupled workloads. Sub-microsecond latency.
AWS ParallelCluster	HPC cluster management	Open-source, auto-scales compute nodes, supports Slurm
FSx for Lustre	High-performance file system	Sub-millisecond latency, 100s GB/s throughput, S3 integration
AWS Batch	Job scheduling	Managed batch processing, supports multi-node parallel jobs

Analogy: HPC on AWS is like renting a supercomputer by the hour — cluster placement groups give you the fast interconnects, EFA gives you the low-level network access, and Lustre gives you the parallel filesystem.

🏠 Dedicated Hosts vs. Dedicated Instances

Feature	Dedicated Instances	Dedicated Hosts
Hardware isolation	Yes — your instances only	Yes — entire physical server
Socket/core visibility	No	Yes — see physical cores, sockets
Instance placement control	No	Yes — place on specific host
BYOL (licensing)	Not supported	Supported (Windows Server, SQL, Oracle)
Pricing	Per-instance premium	Per-host (regardless of instances running)
Use case	Compliance isolation	License compliance + compliance isolation

Decision guide: Need to count physical cores for licensing (Oracle, SQL Server)? → Dedicated Host. Just need hardware isolation for compliance? → Dedicated Instance (simpler, lower cost).

⚠️ VMware Cloud on AWSLIFECYCLE

Important lifecycle change: Following the Broadcom acquisition of VMware, AWS is no longer selling new VMware Cloud on AWS subscriptions. Existing customers can continue using the service but should plan migration strategies.

What it was: VMware SDDC running on bare-metal AWS infrastructure
Migration paths: AWS Application Migration Service, native AWS services, or third-party tools
Exam relevance: Still appears on SAP-C02 — understand the concept but know it's being deprecated for new customers

For the exam: Know that VMware Cloud on AWS provided vSphere, vSAN, NSX on dedicated hardware. In practice, recommend native AWS services for new workloads.

🏗️ AWS Outposts

What: AWS infrastructure (compute + storage) physically deployed in your data center
Form factors: Outposts Rack (42U, full rack) or Outposts Server (1U/2U)
Services available: EC2, EBS, S3, ECS, EKS, RDS, EMR (subset of AWS services)
Connected: Requires network connection back to parent AWS Region
Management: Managed by AWS — they monitor, patch, and maintain hardware
Use cases: Low-latency local processing, data residency, local data processing

Think of it as: A piece of an AWS Region living in your server room. Same APIs, same tools, but the hardware is physically under your roof. AWS still manages it remotely.

📂 FSx File Systems Comparison

FSx Type	Protocol	Best For	Performance
FSx for Lustre	Lustre (POSIX)	HPC, ML training, video processing	100s GB/s, sub-ms latency
FSx for Windows	SMB	Windows workloads, .NET apps, SQL Server	Multi-AZ, AD integration
FSx for NetApp ONTAP	NFS, SMB, iSCSI	Multi-protocol, enterprise storage	Snapshots, tiering, cloning
FSx for OpenZFS	NFS	Linux workloads, data compression	Snapshots, low cost

💻 Demo: Placement Groups & EFA

# Create a cluster placement group
aws ec2 create-placement-group --group-name hpc-cluster \
  --strategy cluster

# Launch instance with EFA in placement group
aws ec2 run-instances --instance-type c5n.18xlarge \
  --placement "GroupName=hpc-cluster" \
  --network-interfaces "DeviceIndex=0,InterfaceType=efa,SubnetId=subnet-xxx"

# Describe EFA interfaces
aws ec2 describe-network-interfaces \
  --filters "Name=interface-type,Values=efa"

🧪 Knowledge Check

Q1: A genomics company needs to run tightly-coupled MPI workloads with sub-microsecond inter-node latency. Which combination should they use?

A) Spread placement group + ENI B) Cluster placement group + EFA C) Partition placement group + ENA D) Dedicated Hosts + ENI

B) Cluster placement group + EFA — Cluster placement groups provide low-latency inter-instance networking. EFA enables OS-bypass for MPI/NCCL, achieving sub-microsecond latency required for tightly-coupled HPC.

Q2: A company must run Oracle Database with per-core licensing. What should they use?

A) Dedicated Instances B) Dedicated Hosts C) Spot Instances D) Outposts

B) Dedicated Hosts — Only Dedicated Hosts provide visibility into physical cores and sockets, required for Bring Your Own License (BYOL) compliance with per-core/per-socket licensing like Oracle.

📝 Module 4 Summary

HPC on AWS

Cluster placement groups + EFA + FSx Lustre + ParallelCluster. OS-bypass networking for MPI.

Dedicated Hosts

Physical server visibility for BYOL licensing. Per-host pricing. Core/socket visibility.

Outposts

AWS infrastructure on-premises. Same APIs. AWS-managed hardware. Rack or server form factor.

VMware Cloud

⚠️ No longer selling new subscriptions. Existing only. Migrate to native AWS services.

🔄 AWS Transform: VMware Migration⚠️ REPLACEMENT

VMware Cloud on AWS no longer accepting new customers. AWS Transform is the migration path.

What is AWS Transform? An AI-powered migration service with specialized agents that automate discovery, planning, and execution of VMware-to-AWS migrations. It replaces AWS Migration Hub and orchestrates the entire journey end-to-end.

Migration Steps: VMware → AWS Native

Step 1: Discovery & Assessment

AWS Transform agent connects to vCenter, discovers all VMs, maps dependencies, identifies OS versions and workload types. Generates a migration readiness report.

Step 2: Intelligent Wave Planning

AI agent groups servers into migration waves based on dependencies, criticality, and affinity. Ensures dependent services migrate together. Suggests target EC2 instance types.

Step 3: Network Configuration Conversion

Translates VMware network constructs (port groups, vSwitches, NSX rules) into AWS-native equivalents (VPCs, subnets, security groups, NACLs). Automated via Transform Network Migration APIs.

Step 4: Replication & Testing (MGN)

Deploy replication agents on VMs → continuous block-level replication to AWS. Launch test instances to validate. No downtime during replication. Uses AWS Application Migration Service (MGN) under the hood.

Step 5: Cutover & Validation

Schedule cutover window → final sync → launch production instances → validate → update DNS → decommission VMware VMs. Minimal downtime (minutes).

Step 6: Post-Migration Optimization

Right-size instances with Compute Optimizer. Modernize: containerize workloads (ECS/EKS), adopt managed services (RDS, ElastiCache), implement auto-scaling. Eliminate VMware licensing costs entirely.