Skip to content
Certification Guide

Task 3.4 · 22 min read

High-Performing and Scalable Network Architectures

Services in this lesson

  • AWS Global Acceleratorin scope
  • AWS Transit Gatewayin scope
  • Amazon VPCin scope
  • AWS Direct Connectin scope
  • AWS Site-to-Site VPNin scope
  • AWS PrivateLinkin scope
  • Amazon EC2in scope

What the exam tests

  • Matching the global entry point to the workload. AWS Global Accelerator fits fixed IP addresses, TCP or UDP traffic and fast regional failover. CloudFront fits when the answer is caching HTTP content.
  • Choosing VPC (virtual private cloud) peering or a transit gateway by how many VPCs there are and whether traffic must pass through a hub.
  • Choosing Direct Connect or Site-to-Site VPN from the bandwidth and consistency the stem asks for. The per-tunnel and per-port ceilings decide it.
  • Choosing a gateway endpoint or an interface endpoint for AWS service traffic, based on where the caller sits and how much bandwidth the path needs.
  • Reading the instance network limits in the stem: one flow versus all flows, baseline versus burst.
  • Picking a faster path between instances: ENA Express, a cluster placement group, or EFA (Elastic Fabric Adapter) for tightly coupled computing.

This lesson owns Global Accelerator, transit gateways, VPC peering for scale and latency, the throughput of VPN and Direct Connect, endpoints as paths, and instance network performance.

VPC layout, security groups, network ACLs and NAT security are taught in 1.2 Secure workloads and applications. VPN and Direct Connect encryption, and the WAF and Shield layer, are there too.

CloudFront and load balancer scaling are in 2.1c Scaling, caching and edge. Route 53 routing is in 2.2 High availability and disaster recovery.

Placement groups are in 3.2 Compute performance. The charges for data transfer and NAT gateways are in 4.4 Network cost.

The core concept

Every network question in this task is about one of three paths: users to your application, your networks to each other, or traffic to AWS services and between instances. Each path has a service built for it, and the stem's numbers pick which one.

The numbers to look for are users per Region, number of VPCs, bandwidth and the size of one flow. A flow is one conversation between two endpoints, such as a single TCP (Transmission Control Protocol) connection. UDP (User Datagram Protocol) sends packets without setting up a connection first.

An analogy: roads for a delivery business

Picture a delivery business with a single phone number that never changes. Callers always dial the same number, and the switchboard forwards each call to whichever depot is open. That's Global Accelerator.

Between depots, you can lay a private lane between each pair of neighbors, or build one central ring road everyone joins once. Private lanes are VPC peering, and the ring road is a transit gateway. For your own warehouse, you either rent a private rail line (Direct Connect) or drive armored vans on the public highway (Site-to-Site VPN).

Where it breaks: a busy road never refuses a vehicle, it just gets slower. A VPN tunnel has a hard ceiling per tunnel: 1.25 Gbps (gigabits per second) for a standard tunnel, and 5 Gbps for a Large Bandwidth Tunnel.

Walk Northgate Freight through it

The diagram shows each path. The walkthrough follows Northgate Freight, a logistics company with offices, a data center and several VPCs.

Enlarge
Each path has its own service. The stem's user count, VPC count, bandwidth and flow size pick the one to use.

Ingrid is Northgate Freight's network engineer. Rafael runs a partner warehouse abroad whose firewall only lets traffic out to an approved list of IP addresses.

  1. Rafael's scanners call the tracking API. They connect to one of two static IP addresses from AWS Global Accelerator. These addresses take traffic in at the edge location (an AWS site close to users) nearest each user.
  2. The nearest edge takes the call. The edge accepts the scanner's TCP connection. It then carries the traffic over the AWS global network instead of the public internet.
  3. The accelerator picks a healthy endpoint.
    • A listener accepts connections on a port, here TCP 443.
    • Each endpoint group holds the endpoints in one Region. Ingrid's are Application Load Balancers (ALBs) in two Regions.
    • Health checks send new connections only to healthy endpoints.
  4. One Region fails. New connections move to the other Region. Rafael's allowlist never changes, because the two IP addresses stay the same across Regions.
  5. The tracking VPC needs the billing VPC. Both attach once to a transit gateway, a regional hub. Its route tables decide which attachment can reach which.
    • Peering the two VPCs would also work, but peering can't pass traffic on to a third network.
  6. The data center sends its nightly sync. Ingrid's AWS Direct Connect circuit, a private line into AWS, reaches the same transit gateway.
    • It enters through a Direct Connect gateway and a transit VIF (virtual interface, a logical connection on the Direct Connect link).
    • A Site-to-Site VPN over the internet is the backup.
  7. The app stores shipment photos in Amazon S3. A gateway endpoint adds a route in the VPC's route table, so the traffic stays on the AWS network with nothing to size.
  8. The data center calls another AWS service privately. An interface endpoint gives that service a private IP address inside the VPC. The data center reaches it over Direct Connect.
  9. Two servers copy a large file between them. That's one flow. It is capped at 5 Gbps, whatever the instance type says.
    • A cluster placement group packs instances close together in one Availability Zone (see 3.2). Inside one, a single flow can reach 10 Gbps.
    • ENA Express is a feature of ENA (Elastic Network Adapter), the standard EC2 network interface. It raises one flow to 25 Gbps, up to the instance's total bandwidth.
    • Both instances must support ENA Express and have it turned on.

Steps 1 to 4 were the user path, steps 5 to 6 the path between networks, and steps 7 to 9 the path to services and between instances.

The rules the exam tests

  • Global Accelerator gives two static IP addresses and fast failover for TCP or UDP. It doesn't cache. CloudFront caches HTTP responses.
  • VPC peering is a one-to-one link between two VPCs, with no hub to size. It is not transitive: VPC B can't reach VPC C through VPC A.
  • A transit gateway routes between all its attachments: VPCs, VPNs, Direct Connect gateways and peered transit gateways. Many VPCs connect without a mesh of peerings.
  • Site-to-Site VPN runs IPsec (encrypted tunnels) over the internet, with two tunnels per connection. A standard tunnel carries up to 1.25 Gbps.
  • Direct Connect is a private circuit with consistent latency. Dedicated ports run at 1, 10, 100 or 400 Gbps. It doesn't encrypt traffic on its own.
  • A gateway endpoint serves only S3 and DynamoDB, and only from its own VPC. An interface endpoint also serves on-premises and peered callers.
  • One flow between instances is capped at 5 Gbps. A cluster placement group raises it to 10 Gbps, and ENA Express to 25 Gbps.
Deep diveGlobal Accelerator details

An accelerator has two static IPv4 addresses. A dual-stack accelerator, which serves both IPv4 and IPv6, has four.

Endpoints can be an Application Load Balancer, a Network Load Balancer (NLB), an Amazon EC2 instance or an Elastic IP address. You choose the Regions they live in.

Deep diveVPN throughput ceilings

A Large Bandwidth Tunnel carries up to 5 Gbps. It works only when the VPN is attached to a transit gateway or to AWS Cloud WAN, a managed service for building a global network.

Large Bandwidth Tunnels are not offered in every Region. The current list is in the Site-to-Site VPN documentation, so check it before you design around one.

Going past one tunnel's ceiling needs ECMP (equal-cost multipath), which spreads traffic across tunnels or across several VPN connections. ECMP relies on BGP (Border Gateway Protocol, dynamic routing) to learn equal-cost routes. VPN ECMP support must also be turned on in the transit gateway.

A VPN has two ends. The customer gateway is the AWS resource that stands for your on-premises VPN device. The AWS end is a transit gateway or a virtual private gateway, the AWS-side VPN endpoint of one VPC.

Deep diveDedicated and hosted Direct Connect

A dedicated connection is a physical port at 1, 10, 100 or 400 Gbps.

A hosted connection comes from a Direct Connect partner at 50 Mbps to 25 Gbps. It is policed at its configured rate: traffic above the rate is dropped.

Deep diveEndpoint and instance limits

A gateway endpoint is a route table entry. It points the service's prefix list (the set of IP ranges AWS publishes for that service) at the endpoint. It adds no network hop you have to size.

An interface endpoint uses AWS PrivateLink. It is one elastic network interface (ENI, a virtual network card) with a private IP address in each subnet you pick. Its throughput is 10 Gbps per Availability Zone (AZ), scaling automatically to 100 Gbps.

The bandwidth listed for an instance type is shared by all its flows. Read the stem for whether one flow or many must go fast.

Decision guide

Decision guide
If the question says…Reach for…Because…
Clients must allowlist a fixed IP, the app uses TCP or UDP on a custom port, users are globalAWS Global Accelerator with ALB, NLB or EC2 endpoints in two RegionsTwo static IP addresses stay the same across Regions. Traffic enters at the nearest edge and rides the AWS network
Static web content or API responses are cacheable and users are globalCloudFront (see 2.1c)Edge caches answer repeat requests without the origin. Global Accelerator routes to endpoints and doesn't cache
Failover between Regions must take effect for new connections within seconds, without waiting on DNS cachesGlobal Accelerator endpoint groups with health checksHealth checks are per endpoint. Traffic moves to a healthy endpoint without waiting for a DNS TTL to expire
Three or four VPCs must talk to each other, with no transit neededVPC peering, one connection per pair of VPCsNo hub to size or pay for, and the path is direct. It stops being practical as pairs multiply
Dozens of VPCs, or VPCs plus on-premises must reach each other through one routing pointAWS Transit Gateway with route tables per segmentRouting is transitive through the hub, and each VPC attaches once. Route tables separate environments
Steady, high bandwidth from the data center, with consistent latency, and the stem names Direct Connect speedsDedicated Direct Connect connection, or a LAG of several connectionsDedicated ports run at 1, 10, 100 or 400 Gbps. A LAG bundles several connections into one logical link
A data center needs connectivity quickly, or a small link is enough, over the internetSite-to-Site VPNEach connection has two tunnels, and a standard tunnel tops out at 1.25 Gbps
VPN throughput above 1.25 Gbps per tunnel, and the connection is on a transit gatewayLarge Bandwidth Tunnels, then ECMP across connections if still shortA Large Bandwidth Tunnel reaches 5 Gbps. ECMP adds their bandwidth together: four 5 Gbps tunnels make 20 Gbps
Lower latency to AWS for a VPN, and the VPN is attached to a transit gatewayAccelerated Site-to-Site VPNTunnel entry points sit on Global Accelerator, so the internet path is shorter. Existing connections can't switch
S3 or DynamoDB traffic from private subnets must stay on the AWS network without a NAT pathGateway endpoint, with the endpoint route on the subnet's route tableTraffic goes to the service's prefix list through a route-table target. There's no endpoint interface to size
Other AWS services, or your own service, must be reached from on-premises or from a peered VPCInterface endpoint (AWS PrivateLink) in the subnets that need itIts private IP routes over VPN, Direct Connect, a transit gateway or peering. Gateway endpoints don't
One flow between two instances must move faster than 5 Gbps, in one RegionENA Express, supported and turned on in both instancesIt carries TCP and UDP over SRD at up to 25 Gbps per flow, within the instance's total bandwidth
Tightly coupled HPC or distributed ML nodes need the lowest latency between instancesElastic Fabric Adapter (EFA) on supported instance typesOS-bypass through Libfabric, on top of ENA, cuts per-message overhead between nodes. Placement is in 3.2

Terms in the table:

  • DNS TTL (time to live) is how long a client may cache a DNS answer before asking again.
  • NAT (network address translation) lets private instances reach out through a shared public address. A LAG (link aggregation group) bundles Direct Connect connections.
  • ENA Express adds SRD, an AWS transport protocol that spreads one flow over many network paths.
  • EFA serves HPC (high-performance computing) and ML (machine learning). OS-bypass lets the application talk to the network card without the operating system in the way, through the Libfabric library.

Correct pattern

REFERENCE PATTERN: a global API with fixed client IP addresses. Two Regions sit behind an accelerator, and a hybrid link enters a transit gateway that reaches the VPCs through the hub.

# 1. Accelerator: two static IP addresses, one listener on TCP 443.
aws globalaccelerator create-accelerator --name api-accelerator \
  --ip-address-type IPV4 --enabled

aws globalaccelerator create-listener --accelerator-arn <accelerator-arn> \
  --protocol TCP --port-ranges FromPort=443,ToPort=443

# 2. One endpoint group per Region, each pointing at that Region's ALB.
aws globalaccelerator create-endpoint-group --listener-arn <listener-arn> \
  --endpoint-group-region us-east-1 \
  --endpoint-configurations EndpointId=<alb-arn-us-east-1>,Weight=128,ClientIPPreservationEnabled=true

# 3. Transit gateway: attach each VPC once, then let route tables decide reachability.
aws ec2 create-transit-gateway --description "shared hub" --options DefaultRouteTableAssociation=enable,DefaultRouteTablePropagation=enable
aws ec2 create-transit-gateway-vpc-attachment --transit-gateway-id <tgw-id> \
  --vpc-id <vpc-a-id> --subnet-ids <subnet-a1> <subnet-a2>

The accelerator has two static IPs for clients to allowlist. The endpoint groups keep traffic in the nearest healthy Region. A Regional failure removes that Region's endpoints from the rotation, and new connections go to the other Region without a DNS change.

On the hybrid side, the transit gateway is the single routing point. The Direct Connect gateway or the VPN attaches to the same hub, so the on-premises network reaches each VPC through its attachment.

Each VPC route table sends the other VPCs' CIDR blocks (their IP address ranges) to the transit gateway. The transit gateway route table then decides which attachment answers.

Why it works: the static IPs decouple clients from Regional endpoints. The hub replaces a mesh of peerings with one routing table. Bandwidth then depends on the hybrid link's capacity and the per-attachment limit, not on a single VPN tunnel.

Traps

The MTU (maximum transmission unit) is the largest packet a link carries. Packets above 1500 bytes are called jumbo frames, and two traps below depend on them.

TRAPExam traps
The trapWhy it failsCorrect pattern
Using CloudFront for a client allowlist of fixed IPs on a custom TCP portCloudFront is built to serve and cache HTTP and HTTPS. Its edge addresses aren't a fixed pair you controlGlobal Accelerator, which gives two static IPv4 addresses and forwards TCP and UDP to ALBs, NLBs or EC2 endpoints
Sending traffic from VPC B to VPC C through a peering connection to VPC APeering is not transitive. B has no route through A, so you need B-to-C peering or a transit gatewayA transit gateway with an attachment per VPC, and route table entries that send each VPC's CIDR to the hub
Assuming a 1.25 Gbps VPN can carry 2 Gbps because it has two tunnelsTwo tunnels give redundancy, not extra bandwidth. Without ECMP, one connection is capped near 1.25 GbpsA Large Bandwidth Tunnel on a transit gateway attachment, or ECMP across several dynamically routed connections
Turning on Accelerated Site-to-Site VPN for an existing connection, or on a virtual private gatewayAcceleration is chosen only when a connection is created. It works only with transit gateway attachmentsCreate a new accelerated connection on a transit gateway, move the customer gateway to it, then delete the old connection
Running a bursty hosted Direct Connect connection at full line rateHosted connections are policed: traffic above the configured rate is dropped, so bursts get less throughputSize the hosted connection for the peak, or use a dedicated connection of 1 Gbps or more
Using a gateway endpoint for S3 traffic from the data center over Direct ConnectGateway endpoints serve only their own VPC route tables. Traffic from VPN, Direct Connect or a hub can't use themAn interface endpoint in the VPC subnets that the on-premises network can reach, with private DNS enabled for the service name
Expecting jumbo frames end to end through a Site-to-Site VPNA transit gateway VPN attachment accepts packets up to 1500 bytes, but the usable IPsec tunnel MTU is 1446 (MSS 1406). No jumbo framesSize packets for the 1446-byte tunnel. For bulk transfer, use Direct Connect: a private VIF supports 9001 bytes and a transit VIF 8500
Expecting one TCP flow between two instances to reach 10 Gbps because the instance type says 10 GbpsOne flow is capped at 5 Gbps, or 10 Gbps in a cluster placement group. Rated bandwidth is shared by all flowsENA Express on both supported instances for up to 25 Gbps in one Region, or a cluster placement group for up to 10 Gbps
Sending 9001-byte jumbo frames to an interface endpointInterface endpoints take 8500 bytes. Larger packets drop, and no ICMP message tells the sender to shrink themKeep packets at or under 8500 bytes toward the endpoint, and rely on the endpoint's TCP MSS clamping for normal TCP traffic

ICMP is the protocol for network error messages. TCP MSS (maximum segment size) is the largest data chunk one TCP packet carries. VPC endpoints clamp it, so normal TCP traffic fits within 8500 bytes.

Compare

Compare
AspectVPC peeringAWS Transit Gateway
ShapeOne-to-one link between two VPCsRegional hub with an attachment per VPC, VPN, Direct Connect gateway or peered transit gateway
Transitive routingNo, traffic can't cross a middle VPCYes, through the hub's route tables
ScalePractical for a few pairsUp to 5,000 attachments per transit gateway (adjustable)
Bandwidth per linkNot capped by a hub; limited by the instancesUp to 100 Gbps per VPC attachment per AZ, each direction
MTU9001 within a Region, 8500 between Regions8500 between VPCs, Direct Connect and peering attachments. VPN: 1500 at the attachment, 1446 in the tunnel
Reach for it whenTwo or three VPCs need a direct, cheap pathMany VPCs, segmented routing or hybrid connections
Compare
AspectSite-to-Site VPN, standard tunnelSite-to-Site VPN, Large Bandwidth TunnelAWS Direct Connect
PathInternet, IPsec, two tunnelsInternet, IPsec, two tunnelsPrivate circuit to a Direct Connect location
Bandwidth per unitUp to 1.25 Gbps per tunnelUp to 5 Gbps per tunnelDedicated 1, 10, 100 or 400 Gbps; hosted 50 Mbps to 25 Gbps
Scaling past one unitECMP on a transit gateway, dynamic routingECMP across connections, for example 20 Gbps from two connectionsLAG of matching connections
Attachment neededVirtual private gateway or transit gatewayTransit gateway or Cloud WAN onlyDirect Connect gateway, private or transit VIF
Encryption by defaultYes, IPsecYes, IPsecNo, add MACsec or a VPN over the link (see 1.2)
Reach for it whenFast setup, modest bandwidthHigher bandwidth on a VPN, or Direct Connect backupSteady, high bandwidth and consistent latency

MACsec is link-layer encryption between your router and the Direct Connect device.

Compare
AspectGateway endpointInterface endpoint (AWS PrivateLink)
ServicesAmazon S3 and DynamoDB onlyMost AWS services, and your own services behind an NLB
How it is reachedRoute table entry to the service's prefix listElastic network interface with a private IP in each chosen subnet
Reachable from VPN, Direct Connect or a transit gatewayNoYes
Reachable from a peered VPCNo: a peered VPC can't use another VPC's gateway endpointYes, when the peer routes to the endpoint's subnets
ThroughputNo endpoint interface to size; the route table carries it10 Gbps per AZ by default, scaling automatically to 100 Gbps
Reach for it whenPrivate S3 or DynamoDB access from the VPC itselfOn-premises or peered callers, or any service without a gateway endpoint

Worked exam scenario

Exam scenario

A payments company runs a TCP-based API behind an Application Load Balancer in us-east-1 and in eu-west-1. Partner banks allowlist one fixed pair of IP addresses, and the API must keep serving from the other Region within seconds if one Region fails. The team wants the LEAST operational overhead. What should a solutions architect do?

  1. A.Create a Route 53 latency-based record set for the API name, with a low TTL.
  2. B.Put Amazon CloudFront in front of both ALBs and allowlist the CloudFront IP ranges.
  3. C.Create an AWS Global Accelerator accelerator with a TCP listener and one endpoint group per Region, each pointing at its ALB.
  4. D.Create a Site-to-Site VPN from each partner bank to each Region's VPC.
Show the walkthrough
  1. Highlight the keywords. Fixed IP addresses, TCP, failover within seconds, and LEAST operational overhead. The partners need one pair of addresses that never changes.
  2. Eliminate A. Latency-based DNS can route users to the closest Region. But the addresses change when a Region's records change, and failover waits on DNS caching even with a low TTL.
  3. Eliminate B. CloudFront is built for HTTP and HTTPS responses and caches them. It doesn't give you two fixed IP addresses you control, so the partners would have to allowlist CloudFront's broad edge ranges.
  4. Eliminate D. A VPN creates a private link between a bank and AWS, but these are public clients on the internet. A VPN per bank and per Region is also far more operational work than allowed.
  5. Land on C. Global Accelerator gives two static IPv4 addresses. The nearest edge accepts each TCP connection, and health checks send new connections to the other Region within seconds, without DNS changes.

Answer: C

Check your understanding

Question 1: A company has six VPCs in one Region. Each VPC must reach every other VPC, and a shared-services VPC must also reach them all. Which design gives the simplest routing with the most room to grow?
Question 2: A hybrid application needs 15 Gbps from an on-premises data center to a VPC. The stem allows an internet path and the team wants the transit gateway. Which TWO actions meet the requirement? (Choose TWO.) (Choose TWO.)
Question 3: A team peers VPC A with VPC B. VPC A has a DynamoDB gateway endpoint, and instances in VPC B's private subnets must reach DynamoDB without a NAT gateway. What should the team do?
Question 4: Two instances in the same Availability Zone must exchange one TCP flow of 15 Gbps. Both are large supported instances. Which option meets the requirement?

Build it

Do the network performance lab. You build three VPCs joined by peering and watch Reachability Analyzer refuse a path that needs a transitive hop. Then you replace the peering routes with a transit gateway and see the same path become reachable.

Key takeaways

  • Global users who need fixed IP addresses, TCP or UDP, and fast regional failover get AWS Global Accelerator. Cacheable HTTP content gets CloudFront.
  • VPC peering is direct and one-to-one, and it is not transitive. A transit gateway is the hub for many VPCs and for hybrid links.
  • A Site-to-Site VPN tunnel carries 1.25 Gbps, or 5 Gbps as a Large Bandwidth Tunnel on a transit gateway. Direct Connect is the dedicated path, at 1 to 400 Gbps per port.
  • A gateway endpoint covers S3 and DynamoDB from the VPC's route tables. An interface endpoint covers other services, and callers can reach it over VPN, Direct Connect, a transit gateway or peering. Either keeps AWS service traffic on a short private path.
  • A single flow is capped at 5 Gbps, or 10 Gbps inside a cluster placement group. ENA Express raises it to 25 Gbps within a Region, and EFA serves tightly coupled HPC and ML nodes.
  • Pick the network path by where traffic starts and ends: Global Accelerator for fixed IPs to global users, a transit gateway for many VPCs, Direct Connect for steady high bandwidth.

Hands-on lab: Peering Is Not Transitive

Sources

Last reviewed