AWS Advanced Networking Specialty (ANS-C01): Optimization & Efficiency

This cram sheet focuses on Task 3.3: Optimize AWS networks for performance, reliability, and cost-effectiveness. It distills critical decision-making criteria for VPC connectivity, high-performance computing, and DNS strategies.

Topic Weighting

Key Concepts Summary

1. Connectivity: Peering vs. Transit Gateway (TGW)

• VPC Peering: Direct connection. Lowest latency, lowest cost (no hourly processing fee, only data transfer). Best for 1-to-1 or small-scale within a region.
• Transit Gateway: Hub-and-spoke. Simplifies management for 100s of VPCs/Accounts. Higher cost (hourly attachment fee + data processing fee) but supports transitive routing and Inter-Region peering.

2. Network Interfaces

• ENA (Elastic Network Adapter): Standard for high performance (up to 100 Gbps). Uses SR-IOV.
• EFA (Elastic Fabric Adapter): Critical for HPC/ML. Provides OS bypass to reduce latency and jitter. Required for MPI/NCCL workloads.

3. Maximum Transmission Unit (MTU)

• Standard MTU: 1500 bytes (Used for Internet traffic and Inter-Region Peering).
• Jumbo Frames: 9001 bytes (Supported within a VPC and for Intra-Region Peering/TGW). Improves throughput by reducing packet overhead.

4. Route 53 Optimization

• Latency-Based Routing: Routes to the region with the lowest round-trip time.
• Weighted Routing: Useful for Blue/Green deployments and canary testing.
• Health Checks: Prerequisite for DNS Failover to ensure reliability.

Visual Anchors

Decision Logic: VPC Interconnectivity

Frame Size Comparison (MTU)

Common Pitfalls

[!WARNING] The MTU Trap: MTU mismatch (e.g., 9001 inside VPC trying to reach 1500 over VPN) leads to packet loss or fragmentation. Always ensure Path MTU Discovery (PMTUD) is functioning (ICMP Type 3 Code 4 must not be blocked).

• Cost Oversight: Forgetting that Transit Gateway charges per GB of data processed in addition to hourly fees. VPC Peering has no processing fee.
• EFA Limitation: EFA is not supported on all instance types and cannot cross subnets unless using specific cluster placement groups.
• Subnet Exhaustion: Using small CIDRs (/28) in Auto Scaling groups, leading to no available IP addresses errors during scale-out events.

Mnemonics / Memory Triggers

• P.E.E.R: Performance, Economy, Easy Routing (Direct).
• E.F.A: Extremely Fast Apps (HPC/OS Bypass).
• R53 - LAW: Latency, Availability (Health Checks), Weighted (Traffic split).

Formula / Equation Sheet

Practice Set

1. Scenario: A company has 50 VPCs in a single region that all need to share a centralized logging service. Which connectivity option is most cost-effective and manageable?
   • Answer: Transit Gateway. While peering is cheaper per-bit, managing 1,225 peering connections (full mesh) is unfeasible.
2. Scenario: An application requires sub-millisecond latency for MPI-based tightly coupled workloads. Which interface is required?
   • Answer: EFA (Elastic Fabric Adapter).
3. Scenario: You observe high packet loss for traffic going from your VPC to an on-premises data center via Direct Connect. What should you check first?
   • Answer: MTU Mismatch / PMTUD. Ensure the MTU is set to 1500 if the path does not support Jumbo frames.
4. Scenario: You need to route users to the closest application endpoint globally. Which Route 53 policy applies?
   • Answer: Latency-Based Routing.
5. Scenario: How can you resolve IP exhaustion in a VPC without creating a new VPC?
   • Answer: Add a Secondary CIDR block to the VPC.

Fact Recall Blanks

1. VPC Peering does NOT support ________ routing. (Answer: Transitive)
2. The maximum MTU supported over the Internet is ________ bytes. (Answer: 1500)
3. To enable EFA, you must use a ________ placement group. (Answer: Cluster)
4. Route 53 ________ record sets allow you to shift traffic to a new version of an app based on a percentage. (Answer: Weighted)
5. Transit Gateway provides a ________ architecture for VPCs. (Answer: Hub-and-Spoke)

Worked Examples

Example 1: Calculating Connectivity ROI

Problem: You transfer 100 TB of data per month between two VPCs in the same region.

• Option A (Peering): $0/hr attachment, $0/GB processing.
• Option B (TGW): $0.05/hr attachment ($36/mo) + $0.02/GB processing.

Calculation:

• TGW Cost = $36 (Attachment) + (100,000 GB * $0.02) = $2,036/mo.
• Peering Cost = $0/mo (excluding standard inter-AZ data transfer fees).
• Result: Use VPC Peering to save ~$2,000/month if management overhead is acceptable.

Example 2: Optimizing for Throughput

Problem: A video processing cluster in a VPC is capped at 10 Gbps. You need 25 Gbps.

• Step 1: Verify the instance type supports higher bandwidth (e.g., C5n or C6g).
• Step 2: Ensure ENA is installed and enabled.
• Step 3: Move instances into a Cluster Placement Group to minimize physical distance.
• Step 4: Increase the MTU to 9001 (Jumbo Frames) on the OS level to reduce CPU interrupt overhead.
• Result: The overhead reduction and physical proximity allow the cluster to reach the target 25 Gbps.