AWS acceleration: pick the right service for the slow leg
On this page
- Situation and definition
- Simple
- Engineer
- Architect
- Picture it as
- Mechanism: where the request goes
- Try it: choose the path for a request
- Six services and what to add
- Deep dive · what knobs do CloudFront and Global Accelerator have?
- Two pairs often confused
- Common exam traps
- Keyword in the question → service
- Self check
- Question 1: static IPs and a UDP game
- Question 2: cross-continent uploads that keep dropping
- Question 3: hammered DynamoDB reads
- Question 4: flaky VPN, needed within days
- Transfer exercise
- Recall
- Core terms
- Closing and golden rules
Acceleration on AWS does not make a server faster. It shortens the public-internet leg — or skips the trip to the origin entirely with a cache. So choosing a service starts with one question: which leg is slow.
| Service | Leg it accelerates | Protocol | Cache |
|---|---|---|---|
| CloudFront | User ↔ HTTP content | HTTP, HTTPS, WebSocket | Yes, at the edge |
| Global Accelerator | User ↔ TCP/UDP application | TCP, UDP | No |
| S3 Transfer Acceleration | Distant client ↔ S3 bucket | HTTP(S) | No |
| Accelerated VPN | Office ↔ VPC | IPsec VPN over GA infrastructure | No |
| DAX | Application ↔ DynamoDB | DynamoDB API | Yes, in-memory |
| ENA · EFA | EC2 ↔ EC2 | Ethernet, RDMA | No |
Situation and definition
Acceleration on AWS means bringing the request into the AWS private network at the edge location nearest the user, or answering from cache, so it avoids the long and unpredictable public-internet path.
Situation: LotusPlay (a fictional company) runs games and video in us-east-1, with a standby in eu-west-1. Users in Vietnam, Brazil and Germany complain about slowness. Six needs: video streaming, a UDP game that needs static IPs for a partner allowlist, a creator uploading 5 GB files, a DynamoDB leaderboard hammered by reads, a Ho Chi Minh City office whose VPN keeps dropping, and a slow ML training cluster. Each need fits a different service. (Illustrative scenario.)
An edge location is an AWS point of presence in many cities, outside the Regions.
Simple
Instead of driving through village roads with many junctions to a warehouse in the US, you take a private highway from the top of your street. Whatever many people need is stocked in a warehouse nearby.
Engineer
DNS (CloudFront) or an anycast IP (Global Accelerator) takes the user into the nearest edge. From there, traffic rides the AWS backbone to the Region; CloudFront also caches HTTP content so it need not go back to the origin.
Architect
Acceleration does not make the server process faster; it shortens the public-internet leg or skips the trip to the origin. If users are already next to the Region, or the bottleneck sits in the database, the benefit is close to zero.
Source: What is CloudFront, What is Global Accelerator.
Picture it as
Village roads with many junctions stand for the public internet: every transit network can jam, and nobody owns the whole route. A highway on-ramp at the top of the street stands for entering the edge early and riding the AWS backbone. A provincial warehouse stands for the CloudFront cache: popular goods are kept nearby, so customers do not wait for delivery from the central warehouse.
| Real life | AWS |
|---|---|
| Village roads with many junctions | The public internet, many transit networks |
| Highway on-ramp at the top of the street | Edge + AWS backbone: Global Accelerator, S3TA, Accelerated VPN |
| Provincial warehouse | CloudFront cache: a copy of HTTP content held at the edge |
Mechanism: where the request goes
Predict before you look: should a UDP game use CloudFront or Global Accelerator? Where is the second viewer of a video answered from?
UDP game: Global Accelerator, because CloudFront does not take UDP. Second viewer with CloudFront: answered from the edge (cache hit). With Global Accelerator: still from us-east-1, because GA does not cache.
No acceleration
Player in HCMC
Public internet: many hops
us-east-1
Every hop can drop or delay packets; nothing is answered from cache.
CloudFront
Viewer
Nearest edge: cache
AWS backbone
Origin us-east-1
Next time the HTTP content is served right at the edge — step 4 — and does not travel to us-east-1.
Global Accelerator
UDP player
2 anycast IPs at the edge
AWS backbone
Endpoint group us-east-1
On failure: endpoint group eu-west-1
A health check sees us-east-1 fail and traffic shifts to eu-west-1 while the 2 static IPs stay the same — step 5. GA does not cache.
Global Accelerator only
5 / 5
- No acceleration: packets cross many transit networks all the way to us-east-1, and every hop is a place where they can be dropped or delayed.
- With acceleration: DNS (CloudFront) or an anycast IP (Global Accelerator) takes the player into the nearest edge; the public-internet leg becomes very short.
- From the edge, the request rides AWS private backbone to the Region — this is the part Global Accelerator, S3TA and Accelerated VPN accelerate.
- CloudFront only: HTTP content is kept at the edge, so later viewers are served right there and never go to us-east-1.
- Global Accelerator only: a health check sees us-east-1 fail, traffic shifts to eu-west-1 while the 2 static IPs do not change, with no DNS wait.
Both diagrams are models of the path, not a description of AWS internal infrastructure, and they do not measure time — the bar segments are equal. The frame highlighted in the primary colour marks each service's branch point and the backbone step.
Try it: choose the path for a request
The source page has a click-through lab; this post replaces it with predict-then-check. LotusPlay's scenarios: watch video, play a UDP game, upload 5 GB to S3, us-east-1 outage; services: no acceleration, CloudFront, Global Accelerator, S3TA; plus a choice of first viewer or later viewer.
| Scenario | No acceleration | CloudFront | Global Accelerator | S3TA |
|---|---|---|---|---|
| Watch video | Slow: all the way to us-east-1 | Best fit: later viewers get it from the edge cache | Works, but no cache | Not supported: origin is an ALB |
| Play a UDP game | Slow, packet loss | UDP not supported | Best fit: 2 static IPs, backbone | Not supported |
| Upload 5 GB | Slow, prone to dropping | Can be configured, wrong tool | Not supported: endpoint cannot be S3 | Best fit: add multipart |
| us-east-1 outage | Errors until DNS changes | Fails over GET, HEAD, OPTIONS only | Best fit: health check, IPs unchanged | Not supported |
The model is discrete, built to teach the choice, not to measure latency: it only says which leg crosses the public internet, which leg rides the AWS network, who answers, and whether the service fits. "Later viewer" only makes a difference when there is a cache.
Six services and what to add
| Service | Leg it accelerates | Required · recommended companions | LotusPlay need |
|---|---|---|---|
| CloudFront | User ↔ HTTP content | OAC for S3; TTL, invalidation; signed URLs; WAF | Streaming video, images, web pages |
| Global Accelerator | User ↔ TCP/UDP application | Endpoint group per Region; health check; traffic dial | UDP game, IP allowlist, failover |
| S3 Transfer Acceleration | Distant client ↔ bucket | Multipart; bucket name without dots; dedicated endpoint | Creator uploading 5 GB |
| Accelerated VPN | Office ↔ VPC | Transit Gateway (cannot be used with a VGW) | Flaky HCMC VPN |
| DAX | Application ↔ DynamoDB | Runs in a VPC; read acceleration only | Read-heavy leaderboard |
| ENA · EFA | EC2 ↔ EC2 | Supported instances; cluster placement group | ML training cluster |
Deep dive · what knobs do CloudFront and Global Accelerator have?
CloudFront: OAC (blocks direct S3 access), signed URLs/cookies, CloudFront Functions (light JS at the viewer) and Lambda@Edge (heavier logic, origin side too), origin groups, geo restriction.
Global Accelerator: TCP/UDP listeners, endpoint groups per Region, traffic dial (percentage into a Region), endpoint weights, client affinity, custom routing (maps players to the right EC2 instance and port).
Source: Global Accelerator components, S3 Transfer Acceleration.
Two pairs often confused
| Criterion | CloudFront | Global Accelerator |
|---|---|---|
| Cache | Yes, served from the edge | No, every packet reaches an endpoint |
| Protocol | HTTP, HTTPS, WebSocket | TCP, UDP |
| Entry point | A *.cloudfront.net domain name | 2 static anycast IPs |
| Target | S3, ALB, EC2, any HTTP origin | ALB, NLB, EC2, Elastic IP |
| Failover | Origin group, GET, HEAD, OPTIONS only | Health check, no DNS TTL wait |
| Criterion | DAX | ElastiCache |
|---|---|---|
| Made for | DynamoDB only | Any source: RDS, APIs, sessions |
| Code changes | Minimal, DynamoDB-compatible API | You write the read/write cache logic |
| Writes | Write-through, no faster | Depends on the strategy you pick |
| Consistency | Eventually consistent; strong reads go straight to the table | Depends on how you invalidate |
Source: CloudFront origin failover, DAX and consistency.
Common exam traps
| The question says | Reality |
|---|---|
| Global Accelerator caches like CloudFront | No cache; it only moves packets onto the backbone |
| CloudFront works for a UDP game | CloudFront is HTTP(S) and WebSocket only |
| Route 53 latency routing fails over as fast as GA | DNS is cached per TTL; GA switches the path immediately |
| S3TA lets millions read the same file | That is CloudFront's job |
| Turn on Accelerated VPN on a Virtual Private Gateway | Only VPNs attached to a Transit Gateway support it |
| Attach an Elastic IP to an ALB for static IPs | ALB cannot take an EIP; put Global Accelerator in front |
Source: Accelerated Site-to-Site VPN, CloudFront WebSocket support.
Keyword in the question → service
| The question says | Pick |
|---|---|
| Static IPs, TCP/UDP, multi-Region failover | Global Accelerator |
| Static content, video for a global audience | CloudFront (+ OAC if the origin is S3) |
| Uploads from many continents into one bucket | S3 Transfer Acceleration + multipart |
| DynamoDB reads in microseconds, minimal code changes | DAX |
| A distant branch VPN that keeps dropping | Accelerated Site-to-Site VPN on a Transit Gateway |
| HPC, MPI needing the lowest latency between nodes | EFA + cluster placement group |
Self check
Question 1: static IPs and a UDP game
A LotusPlay partner opens its firewall to only 2 fixed IPs. The game uses UDP, with servers in us-east-1 and eu-west-1. What do you pick?
- CloudFront in front of the game servers.
- Global Accelerator with an endpoint group in both Regions.
- An Elastic IP on each EC2 instance.
Answer yourself before reading the explanation
- CloudFront — Not quite. It takes HTTP, HTTPS and WebSocket only, and does not give you 2 fixed IPs to allowlist.
- Global Accelerator — Correct. 2 static anycast IPs, UDP support, and routing to endpoint groups in both Regions by health check.
- Elastic IP — Not quite. An Elastic IP attaches to one instance in one Region; it cannot fail over to another Region while keeping the IP.
Question 2: cross-continent uploads that keep dropping
Creators in Brazil and Germany upload 5 GB videos into a us-east-1 bucket and transfers often drop midway. Best approach?
- Put CloudFront in front of the bucket for caching.
- Turn on S3 Transfer Acceleration and use multipart upload.
- Put Global Accelerator in front of the bucket.
Answer yourself before reading the explanation
- CloudFront — Not quite. CloudFront shines at caching content for repeat reads; uploads are not cached, and S3 has a feature built for this.
- S3TA + multipart — Correct. S3TA brings the upload into an edge near the creator and then onto the backbone; multipart sends parts in parallel and retries only the failed part.
- Global Accelerator — Not quite. GA endpoints are ALB, NLB, EC2 and Elastic IP — it cannot point at an S3 bucket.
Question 3: hammered DynamoDB reads
A DynamoDB leaderboard gets hundreds of thousands of reads per minute, needs microseconds, and the dev team does not want to rewrite cache logic. What do you pick?
- DAX.
- ElastiCache for Redis.
- CloudFront.
Answer yourself before reading the explanation
- DAX — Correct. DAX is an in-memory cache made for DynamoDB, with a compatible API, so you only swap the client.
- ElastiCache — Not quite. It can work, but you must write the read, write and invalidation logic yourself — against the requirement of minimal code changes.
- CloudFront — Not quite. CloudFront caches HTTP responses at the edge for end users; it does not speed up a VPC application's DynamoDB reads.
Question 4: flaky VPN, needed within days
A VPN from the Ho Chi Minh City office to the us-east-1 VPC keeps dropping. The VPN currently attaches to a Virtual Private Gateway. Acceleration is needed within days.
- Turn on acceleration on the current VPN.
- Create a new VPN attached to a Transit Gateway and enable acceleration.
- Move to Direct Connect.
Answer yourself before reading the explanation
- Enable on the current VPN — Not quite. Accelerated VPN does not support a Virtual Private Gateway, and acceleration cannot be toggled on an existing connection.
- New VPN on a Transit Gateway — Correct. Acceleration exists only for Site-to-Site VPNs attached to a Transit Gateway; create the new VPN on a TGW and enable acceleration.
- Direct Connect — Not quite. Direct Connect is stable but usually takes weeks to install, so it does not fit days.
Transfer exercise
50,000 IoT devices send MQTT over TCP to NLBs in ap-southeast-1 and eu-central-1. The partner needs a static-IP allowlist; when a Region fails, connections must move to the other one immediately. Design it.
Answer: Global Accelerator, a TCP listener, two endpoint groups (one NLB each), with health checks enabled.
Grading criteria: (1) rule out CloudFront because MQTT is not HTTP; (2) mention the 2 static anycast IPs for the allowlist; (3) fail over by health check, not DNS TTL; (4) no cache is needed. Missing (1) or (2) is not a pass.
Recall
Why is the second viewer of a video faster with CloudFront but not with Global Accelerator?
CloudFront keeps a copy of the HTTP content at the edge, so the later viewer is served right there. Global Accelerator only shortens the path; with no cache, every request still reaches us-east-1.
What are the three questions for choosing an acceleration service?
(1) Which leg is slow: user ↔ application, client ↔ S3, office ↔ VPC, application ↔ DynamoDB, or EC2 ↔ EC2? (2) Can it be cached? (3) What protocol, and are static IPs required?
Core terms
| Term | Meaning |
|---|---|
| Edge location | An AWS point of presence in many cities, outside the Regions |
| Anycast | Many locations advertising the same IP; packets reach the nearest one |
| AWS backbone | AWS's private network linking edge locations to the Regions |
| Cache hit | The content is already at the edge or in the cache, so no origin trip is needed |
| Origin group | A pair of origins for failover; CloudFront fails over GET, HEAD, OPTIONS only |
| OAC | Origin Access Control: blocks direct access to an S3 bucket |
| Traffic dial | The percentage of traffic GA sends into an endpoint group |
| Write-through | Writes go through the cache to the source before success is returned, as in DAX |
| Cluster placement group | Packs instances close together in one AZ for the lowest latency |
Closing and golden rules
Acceleration is not about making the server faster. It is about shortening the path, or skipping the trip.
- Choose by the slow leg, not by service name.
- Caching is not the same as a faster route: CloudFront and DAX answer in place; Global Accelerator, S3TA and Accelerated VPN only change the route.
- The required companions decide success: multipart, Transit Gateway, OAC, cluster placement group.
- When users are next to the Region, or the bottleneck is the database, route acceleration is close to useless.
Enter the edge early · Ride the backbone · Cache when you can
Source: What is CloudFront · CloudFront origin failover · CloudFront WebSocket · What is Global Accelerator · Global Accelerator components · S3 Transfer Acceleration · S3 pricing — no charge when not faster · Accelerated Site-to-Site VPN · DAX · DAX and consistency · ENA · EFA · Placement groups. Checked on 2026-10-08.
Related posts
AWS IAM policies: follow the request, find the Deny
Animated AWS IAM walkthroughs explain policy types, permissions boundaries, KMS, cross-account access and why one matching Deny overrides every Allow.
Speeding up EC2 and Lambda boot time
How to shorten Lambda cold starts and EC2 scale-out time with SnapStart, Provisioned Concurrency and warm pools, plus where each one stops working.
Attach AWS WAF in the right place and close the bypass
Where AWS WAF can attach, how ALB differs from NLB, body inspection limits and locking the origin, with six scenarios and config examples.
