Task 3.1 · 22 min read
High-Performing and Scalable Storage
Services in this lesson
- Amazon S3in scope
- Amazon EBSin scope
- Amazon EFSin scope
- Amazon FSx (for all types)in scope
- AWS Storage Gatewayin scope
- AWS DataSyncin scope
What the exam tests
- Matching a workload to object, file or block storage from how it reads and writes data, before thinking about any specific feature.
- Reading the performance number in the stem and picking the setup that meets it. The number may be requests per second, IOPS (input/output operations per second), MiB/s (mebibytes per second) or latency.
- Knowing the tuning levers: S3 prefixes and multipart upload, an EBS volume type, an EFS throughput mode, or an FSx file system.
- Spotting which storage scales on its own (S3, EFS) and which needs you to provision or modify capacity (EBS, most FSx file systems).
- Picking the hybrid pattern when applications stay on premises: a Storage Gateway type for ongoing access, AWS DataSync for moving data.
This lesson owns performance and scale. Choosing S3 storage classes and lifecycle rules to cut cost is taught in 4.1 Cost-optimized storage.
The core concept
Every storage question starts with one fact: how does the application reach its data? Pick object, file or block storage from that answer, then tune the chosen service to the number in the stem.
An analogy: a film studio's storage
A cloakroom hands back a whole item to anyone with the ticket. A shared filing cabinet lets every editor open folders by name. A personal desk drawer belongs to one person, who arranges it as they like.
The cloakroom is object storage, the cabinet is file storage, and the drawer is block storage. Where it breaks: S3 lets any number of clients read the same object, and can return just part of it. The drawer isn't strictly single-user either, because EBS Multi-Attach shares one io1 or io2 volume with up to 16 instances.
Walk one studio through it
The diagram is the decision flow. The walkthrough follows one company down each branch.
Lumen Reel is a video editing studio. Farah is its cloud architect, and Bruno runs the editing office. Farah answers the diagram's first question, "How is the data reached?", once for each workload.
- Clients upload raw footage. Upload tools call an HTTP API, and any number of them read the same file. That is object storage: Amazon S3 (Simple Storage Service).
- S3 stores each file as an object under a key, its full name in the bucket, the S3 container that holds objects.
- A prefix is the start of a key, like a folder path. S3 scales request rate per prefix.
- A few Linux edit servers share project files. Each one mounts the same directory tree (attaches it like a local folder) and opens files by path. That is file storage.
- Amazon EFS (Elastic File System) serves Linux clients over NFS (Network File System). It grows and shrinks with the files.
- A 400-server render farm needing hundreds of GB/s would point to FSx for Lustre instead, as in the worked scenario.
- If Bruno's editors used Windows, Farah would pick an Amazon FSx file system instead. FSx offers Windows, Lustre, NetApp ONTAP or OpenZFS engines.
- The asset catalog database runs on one server. That Amazon EC2 instance (virtual server) needs a raw disk its operating system formats. That is block storage.
- The data must survive a stop, so Farah picks Amazon EBS (Elastic Block Store), a disk reached over the network.
- The volume type and the instance both cap its speed, so she sizes both.
- The transcoding scratch space can be rebuilt. So the diagram's "No" branch fits: instance store, disks physically attached to the host.
- Bruno's office keeps its own servers. The diagram's last branch is AWS Storage Gateway. It gives on-premises servers a local file, block or tape interface, with a local cache. File and tape data end up in AWS, and cached volumes keep primary data in AWS. Stored volumes keep it on premises.
- Bruno's old archive must move to AWS once. That is a transfer job for AWS DataSync, not a gateway.
Only after the type is chosen do the stem's numbers matter. Farah then tunes each service to its own number.
The rules the exam tests
- Access pattern first. HTTP API means S3. A shared mount means EFS or FSx. One server's raw disk means EBS or instance store.
- S3 scales by design. Request rate grows per partitioned prefix. Throughput grows with parallel connections, such as multipart uploads and byte-range requests (fetching parts of an object).
- EBS gives exactly what you provision. The volume type and the instance both set the ceiling.
- EFS scales on its own. Capacity follows files. Elastic throughput follows traffic, and Provisioned throughput is a rate you set. General Purpose is the default performance mode.
- FSx is for a protocol or feature set that rules EFS out.
- Hybrid splits two ways. Storage Gateway is for ongoing access. DataSync moves data, covered in 3.5 Data ingestion.
Deep diveStorage Gateway interfaces
AWS Storage Gateway gives on-premises servers a local NFS, SMB, iSCSI or tape interface. NFS and SMB carry files, and iSCSI carries block volumes.
The gateway keeps a local cache, so recently used data stays close to the applications.
Farah's walkthrough named a service for each workload. The next part covers the numbers each service is tuned to.
Units in the stem
- IOPS (input/output operations per second) counts small reads and writes. Databases care about it.
- Throughput, in MiB/s (mebibytes per second), measures data moved. Streaming and analytics care about it.
- Latency is how long one operation takes. First-byte latency is the wait before data starts arriving.
Decision guide
Terms are explained under the table.
| If the question says… | Reach for… | Because… |
|---|---|---|
| S3 returns 503 Slow Down as request rate grows | Spread keys across more prefixes and retry with backoff | Each partitioned prefix supports at least 3,500 writes and 5,500 reads per second |
| Large uploads to S3 are slow or restart after network drops | Multipart upload with parts sent in parallel | AWS encourages it above 100 MB. A single PUT tops out at 5 GB |
| Users on other continents upload to one central bucket | S3 Transfer Acceleration | Uploads enter at the nearest CloudFront edge location, then cross the AWS network |
| Analytics or ML jobs need consistent single-digit millisecond object access from one AZ | S3 Express One Zone in a directory bucket next to the compute | First-byte latency is up to 10x lower than S3 Standard, in one AZ you choose |
| A database on one EC2 instance needs a few thousand to tens of thousands of IOPS | EBS gp3, provisioning IOPS and throughput separately from size | Includes 3,000 IOPS and 125 MiB/s, and scales to 80,000 IOPS and 2,000 MiB/s |
| Sustained sub-millisecond latency or more than 80,000 IOPS on one volume | EBS io2 Block Express on a Nitro instance | Up to 256,000 IOPS and 4,000 MiB/s, with average latency under 500 microseconds |
| Large sequential reads such as logs or a data warehouse on one instance | EBS st1 | Built for streaming, up to 500 MiB/s. It can't be a boot volume |
| Highest local IOPS for a cache or scratch data that can be rebuilt | Instance store | Disks are attached to the host. Data is lost when the instance stops or terminates |
| Many Linux instances across AZs share files and the size is unpredictable | Amazon EFS with Elastic throughput | A Regional NFS file system for many clients. Capacity and throughput scale automatically |
| Windows servers need SMB shares joined to Active Directory | FSx for Windows File Server | Native Windows file system and SMB. EFS doesn't support Windows clients |
| HPC or ML training needs hundreds of GB/s and must read a dataset that lives in S3 | FSx for Lustre linked to the S3 bucket | Sub-millisecond latency at very high throughput, with S3 objects presented as files |
| One file system must serve NFS, SMB and iSCSI clients, or the team runs NetApp today | FSx for NetApp ONTAP | Multiprotocol access, plus ONTAP snapshots, cloning and replication |
| On-premises apps write files over NFS or SMB and the files should land as S3 objects | Amazon S3 File Gateway | File protocols on premises, objects in S3, and a local cache for recent data |
| On-premises apps need low-latency access to the whole dataset, with backups in AWS | Volume Gateway, stored volumes | Primary data stays on local disks. Snapshots go to AWS asynchronously |
| Existing backup software writes to tape | Tape Gateway | A virtual tape library over iSCSI, so the backup software keeps working unchanged |
| Move terabytes of files from an on-premises NAS into EFS, FSx or S3, once or on a schedule | AWS DataSync | A managed transfer service. A gateway is for ongoing access, not migration |
Terms in the table:
- 503 Slow Down is the error S3 returns when requests arrive faster than a prefix can take them. S3 scales up gradually, so 503s can appear while it scales.
- The fix is spreading keys across prefixes and retrying with backoff: waiting longer before each retry. The AWS SDK (software development kit) is the code library apps use to call AWS.
- A PUT is the request that uploads one object.
- Multipart upload splits one object into parts that upload separately and in parallel.
- A CloudFront edge location is an AWS site near users. S3 Express One Zone stores objects in a directory bucket in one Availability Zone (AZ): one or more data centers in a Region.
- EBS volume types: gp3 (General Purpose SSD), io2 (Provisioned IOPS SSD) and st1 (Throughput Optimized HDD).
- Nitro is the AWS-built hardware and hypervisor system that current instance types run on.
- HPC is high-performance computing, and ML is machine learning. A NAS (network-attached storage) is an on-premises file server.
- Active Directory is the Microsoft directory that manages Windows users and computers.
- SMB (Server Message Block) is the Windows file-sharing protocol. iSCSI (Internet Small Computer Systems Interface) sends disk commands over a network.
- A snapshot is a point-in-time backup copy of a volume.
Deep diveVolume Gateway: stored versus cached volumes
Volume Gateway presents iSCSI block volumes to on-premises servers. It comes in two modes.
With stored volumes, primary data stays on local disks and snapshots go to AWS asynchronously. With cached volumes, primary data lives in AWS and only hot data is cached locally.
Correct pattern
REFERENCE PATTERN: a performance-tuned block volume and a shared file system for the same application tier, plus an S3 key layout that spreads load.
Elastic Volumes is the EBS feature that changes a volume's type, size, IOPS or throughput while it stays attached. It works on supported instances, and size can only grow.
# 1. Block: a gp3 data volume for one database instance.
# IOPS and throughput are set independently of size.
aws ec2 create-volume \
--availability-zone <az-of-the-instance> \
--volume-type gp3 --size 200 \
--iops 12000 --throughput 500
# Later, raise performance in place (Elastic Volumes). On current-generation
# instances this usually needs no detach or restart.
aws ec2 modify-volume --volume-id <vol-id> --iops 16000 --throughput 750
# 2. File: one EFS file system shared by every web server in every AZ.
aws efs create-file-system \
--performance-mode generalPurpose \
--throughput-mode elastic \
--encrypted
# Create one mount target per AZ that runs clients, then on each Linux host:
sudo mount -t efs -o tls <file-system-id> /mnt/shared
# 3. Object: spread hot keys across prefixes so each one scales separately.
s3://<bucket>/uploads/2026/11/15/a1/<object-key>
s3://<bucket>/uploads/2026/11/15/b7/<object-key>
s3://<bucket>/uploads/2026/11/15/f3/<object-key>
Why it works:
- The volume lives in the same Availability Zone as its instance. It gets exactly the performance the workload needs.
- The web tier shares one file system, so no server copies files to another.
- A mount target is the NFS endpoint, with an IP address, that you create in each AZ. Clients in that AZ mount through it.
- The bucket's request rate grows with the number of prefixes, instead of hitting one prefix's ceiling.
Traps
| The trap | Why it fails | Correct pattern |
|---|---|---|
| Using EBS Multi-Attach as a shared file system for a web fleet | Works only for io1 and io2, in one AZ, for up to 16 Nitro instances. Needs a cluster-aware file system | Use Amazon EFS (Linux) or FSx (Windows, Lustre, ONTAP, OpenZFS) for files many servers share. |
| Moving to io2 because gp3 "is not fast enough" at 3,000 IOPS | 3,000 is only gp3's included baseline. It provisions up to 80,000 IOPS | Raise gp3 IOPS and throughput with Elastic Volumes. Pick io2 for sub-millisecond latency or more than gp3's limits. |
| Expecting a fast volume to fix a slow instance | The instance's EBS bandwidth also caps throughput and IOPS | Check both the volume limits and the instance type's EBS-optimized bandwidth. |
| Storing the only copy of data on instance store because it is fastest | Instance store data is lost when the instance stops, hibernates or terminates | Keep instance store for caches and scratch. Persist data on EBS, EFS or S3. |
| Choosing EFS for Windows servers that need SMB | EFS is NFS only and does not support Windows instances | FSx for Windows File Server (or FSx for NetApp ONTAP for mixed clients). |
| Picking EFS Max I/O mode "for maximum performance" | Previous generation, with higher per-operation latency | Use General Purpose mode with Elastic throughput. Use Provisioned throughput only for a known steady rate. |
| Using Transfer Acceleration or CloudFront to fix 503 Slow Down errors | Those fix distance and latency, not the per-prefix request rate | Spread requests over more prefixes and let the SDK retry with backoff. |
| Using Storage Gateway to migrate a file server once | A gateway gives ongoing hybrid access through a cache. It isn't a bulk mover | AWS DataSync for the migration. Storage Gateway only if on-premises apps keep reading the data afterward. |
An EBS-optimized instance has dedicated bandwidth for EBS traffic.
Deep diveMulti-Attach and Max I/O in detail
EBS Multi-Attach lets one io1 or io2 volume attach to up to 16 Nitro instances in the same AZ. It needs a cluster-aware file system, one built to coordinate writes from several hosts.
Standard file systems such as XFS and ext4 are not designed for simultaneous access from several instances. That is why Multi-Attach is not a shared file system.
EFS Max I/O is a previous-generation performance mode with higher per-operation latency. It isn't available with Elastic throughput or with One Zone file systems.
Compare
| Aspect | Amazon S3 | Amazon EBS | Amazon EFS | Amazon FSx |
|---|---|---|---|---|
| Storage type | Object | Block | File (NFS) | File (SMB, Lustre, NFS, iSCSI, by type) |
| Accessed by | Any number of clients over HTTPS | One instance (io1/io2 Multi-Attach: up to 16) | Many Linux clients across AZs | Many clients, protocol set by the FSx type |
| Scope | Region (Express One Zone: one AZ) | One Availability Zone | Region (or One Zone) | Single-AZ or Multi-AZ deployments, by type |
| How it scales | Automatically; request rate grows per prefix | You modify size, IOPS, throughput | Automatically, with Elastic throughput | Set capacity; adjustable later on most types |
| Typical latency | Slower first byte; Express One Zone: single-digit ms | gp3 single-digit ms; io2 under 0.5 ms average | Best case 1 ms reads, 2.7 ms writes (Standard) | Depends on the FSx type and storage media |
| Reach for it when | Unstructured data, data lakes, static content | Boot volumes and databases on one instance | Shared Linux content, home directories, containers | Windows shares, HPC, NetApp or ZFS migrations |
Worked exam scenario
Exam scenario
A media company runs a rendering farm of 400 Linux EC2 instances that read a 200 TB library of source assets stored in Amazon S3. Each job reads thousands of files and needs sub-millisecond latency and hundreds of GB/s of aggregate throughput. Finished frames must end up back in the same S3 bucket. Which storage solution meets these requirements with the LEAST operational overhead?
- A.Mount the S3 bucket on every instance with a file system driver and read the assets directly from S3.
- B.Create an Amazon EFS file system with Elastic throughput, copy the assets into it with AWS DataSync, and copy frames back to S3 after each job.
- C.Create an Amazon FSx for Lustre file system linked to the S3 bucket and export finished frames back to the bucket.
- D.Attach an io2 Block Express volume with Multi-Attach to the instances and load the assets onto it.
Show the walkthrough
- Mark the requirements. Many Linux clients need sub-millisecond latency and hundreds of GB/s in total. The data lives in S3, results go back to S3, and the stem wants the least operational overhead.
- A keeps the data in S3. But S3 first-byte latency is far above sub-millisecond. Eliminate.
- B gives a shared Linux file system, but EFS reads take about 1 ms at best. Copying 200 TB in and out for every job also adds the overhead the stem rules out. Eliminate.
- D fails on scale and sharing. Multi-Attach reaches only 16 Nitro instances in one AZ, not 400, and needs a cluster-aware file system. Eliminate.
- C is built for this. FSx for Lustre gives sub-millisecond latency at very high throughput. A data repository association, its link to the bucket, presents the S3 objects as files.
- That association's automatic export policy writes new and changed files back to the bucket. There is no separate copy job to build. Answer: C.
Answer: C
Check your understanding
Build it
Do the storage performance lab. You attach a gp3 volume and watch an attach across Availability Zones fail. Then you raise its IOPS while it stays mounted, and share one EFS file system between instances in two AZs.
Key takeaways
- Object, file or block comes first. The service and its settings come second.
- S3 scales by prefix and parallelism: more prefixes for request rate, multipart and byte-range requests for throughput, Transfer Acceleration for distance, Express One Zone for latency.
- EBS is one AZ and usually one instance. Use gp3 by default, io2 for sub-millisecond or extreme IOPS, and st1 for streaming. The instance's EBS bandwidth caps them all.
- Shared files mean EFS for Linux NFS, or the FSx type that matches the protocol or workload: Windows, Lustre, ONTAP or OpenZFS.
- Hybrid: Storage Gateway for ongoing on-premises access through a cache, DataSync for moving data.
- Choose the storage type from how the data is reached (object API, shared file mount or one-instance disk), then tune it to the stem's number: requests, IOPS, throughput or latency.
Hands-on lab: Block Versus Shared File Storage
Sources
- https://docs.aws.amazon.com/AmazonS3/latest/userguide/optimizing-performance.html
- https://docs.aws.amazon.com/AmazonS3/latest/userguide/optimizing-performance-guidelines.html
- https://docs.aws.amazon.com/AmazonS3/latest/userguide/mpuoverview.html
- https://docs.aws.amazon.com/AmazonS3/latest/userguide/qfacts.html
- https://docs.aws.amazon.com/AmazonS3/latest/userguide/upload-objects.html
- https://docs.aws.amazon.com/AmazonS3/latest/userguide/transfer-acceleration.html
- https://docs.aws.amazon.com/AmazonS3/latest/userguide/directory-bucket-high-performance.html
- https://docs.aws.amazon.com/ebs/latest/userguide/ebs-volume-types.html
- https://docs.aws.amazon.com/ebs/latest/userguide/provisioned-iops.html
- https://docs.aws.amazon.com/ebs/latest/userguide/hdd-vols.html
- https://docs.aws.amazon.com/cli/latest/reference/ec2/modify-volume.html
- https://docs.aws.amazon.com/efs/latest/ug/whatisefs.html
- https://docs.aws.amazon.com/efs/latest/ug/mounting-fs-old.html
- https://docs.aws.amazon.com/ebs/latest/userguide/general-purpose.html
- https://docs.aws.amazon.com/ebs/latest/userguide/ebs-volumes-multi.html
- https://docs.aws.amazon.com/ebs/latest/userguide/ebs-modify-volume.html
- https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ebs-optimized.html
- https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/InstanceStorage.html
- https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instance-store-lifetime.html
- https://docs.aws.amazon.com/efs/latest/ug/performance.html
- https://docs.aws.amazon.com/AmazonS3/latest/userguide/s3-express-performance.html
- https://docs.aws.amazon.com/efs/latest/APIReference/API_CreateFileSystem.html
- https://docs.aws.amazon.com/efs/latest/ug/troubleshooting-efs-mounting.html
- https://docs.aws.amazon.com/fsx/latest/WindowsGuide/what-is.html
- https://docs.aws.amazon.com/fsx/latest/LustreGuide/what-is.html
- https://docs.aws.amazon.com/fsx/latest/LustreGuide/autoexport-data-repo-dra.html
- https://docs.aws.amazon.com/fsx/latest/ONTAPGuide/what-is-fsx-ontap.html
- https://docs.aws.amazon.com/fsx/latest/OpenZFSGuide/what-is-fsx.html
- https://docs.aws.amazon.com/filegateway/latest/files3/what-is-file-s3.html
- https://docs.aws.amazon.com/storagegateway/latest/vgw/StorageGatewayConcepts.html
- https://docs.aws.amazon.com/storagegateway/latest/tgw/GettingStartedAccessTapesVTL.html
- https://docs.aws.amazon.com/datasync/latest/userguide/what-is-datasync.html
- https://docs.aws.amazon.com/efs/latest/ug/how-it-works.html
- https://docs.aws.amazon.com/ec2/latest/instancetypes/ec2-nitro-instances.html
- https://docs.aws.amazon.com/ebs/latest/userguide/what-is-ebs.html
Last reviewed