Lesson 8 - Select High-performing Networking Solutions for a Workload.md

Lesson 8 - Select High-performing Networking Solutions for a Workload

8.1 VPC Performance

• Consolidate resources within a single AZ to minimize latency (< 1ms), but this compromises resilience
• Traffic Egress Fragmentation - Jumbo frames are fragmented to 1500 MTU if they use an IGW for egress from the VPC. This is anti-pattern
• Avoid Public AWS Network. Using an IGW to reach AWS services requires traffic crossing public AWS networks. Instead adding a gw or interface endpoint allows traffic to be proxied directly to aws serves.
• Avoid Inline Gateways. An EC2 based inline gateway can be a router, proxy or NAT. This is an anti-pattern because it will be a bottleneck and single point of failure.

8.2 Single-node Performance

• EC2 instance type with ENI (Elastic Network Interface)

  • Directly attached to EC2
  • Maximum 25Gbps but this depends on the instance type, can be much lower

• EC2 instance type with ENA (Elastic Network Adapter)

  • Instance types that end with "n" support ENA
  • Some are bare metal
  • Some use Nitro virtualization
  • No published maximum but significantly higher than ENI
  • ENA require Enhanced Networking to be enabled on instance

• EC2 instance type with EFA (Elastic Fabric Adapter)

  • 2 largest instance types of each family that end with "n" support EFA
  • MPI (Message Passing Interface) one-way latency as low as 15.5 microseconds
  • Maximum 100Gbps but depends on instance
  • It is ENA with added capabilities, and benefits from cluster placement groups

• EC2 Enhanced Networking

  • Using Software or the guest OS to improve the network performance
  • Uses single root I/O virtualization
  • For ENI, just switch the kernel driver to enable this mode
  • For ENA/EFA, switch the kernel driver then attach the resource
  • Higher bandwidth, higher PPS, lower latency, less jitter

• EC2 Placement Groups

  • Designed to influence EC2 instance placement within the region for performance

  • Placement groups can be used within single AZ or across multiple AZ

  • Three types of Placement Groups:

    1. EC2 Cluster Placement Group

       • Deploy all the instances within a single rack.
       • No blocking, no oversubscription
       • High throughput, low latency
       • Use case: Clusters have majority of traffic between group nodes
       • Not designed for resilience

    2. Partition Placement Group

       • You can deploy your resources into multiple places in the same AZ across partition where partition = set of racks
       • Use case: Distributed + Replicated workloads
       • Can span multiple AZ
       • Designed for performance and resilience

    3. EC2 Spread Placement Group

       • Each instance = separate rack
       • Use case: Critical instances that must be kept separate
       • Designed for guaranteed resilience
       • Not for performance
       • Can span multiple AZ

8.3 Hybrid Network Performance

• Connecting two networks together where one of them is not on AWS

• VPG (Virtual Private Gateway)

  • Connecting corporate network through VPN to a VPC
  • Connection will be encrypted with IPSEC and key exchange
  • Traffic traverses the internet
  • There is not way to guarantee specific throughput or latency

• VPC Peering

  • Guaranteed private connection
  • Within the same region the traffic is not encrypted
  • Encrypted if cross region
  • Better performance than DIY or other custom solution
  • No published latency or throughput metrics

• Direct Connect Gateway

  • This is where you run a fibre WAN into an AWS partner data center and they cross connect you into AWS
  • Guaranteed bw upto 10GBps per connection
  • You can connect into your VPC or into the AWS service API gateways
  • Lower latency than VPG
  • Latency numbers are only published to the closest region

• DIY

  • Do it yourself
  • Full flexibility
  • No guaranteed performance