Skip to content

8 · Storage Performance — Revise

Performance · 24% · Compressed recall. Read the Learn guide first.

Storage shape first

Shape Service Attach
Object S3 Anywhere via API
Block EBS / instance store One instance, one AZ
File EFS / FSx Many instances mount concurrently

S3 classes (all 11 nines durability)

Class Use Min duration
Standard Frequent
Intelligent-Tiering Unknown/changing (no retrieval fee)
Standard-IA Infrequent, fast 30 d
One Zone-IA Infrequent + reproducible (1 AZ) 30 d
Glacier Instant Archive, ms access 90 d
Glacier Flexible Archive, min–hrs 90 d
Glacier Deep Archive Coldest, hrs 180 d

S3 perf: ~3,500 write / ~5,500 read requests per prefix/s (scale via more prefixes); multipart >100 MB (single PUT max 5 GB); object max 5 TB; Transfer Acceleration for far uploads.

EBS volume types

Type Cat Max IOPS Max MB/s Pick for
gp3 SSD 16k→80k 2,000 Default (base 3,000/125, independent scaling)
gp2 SSD 16k 250 Legacy (3 IOPS/GiB)
io2 BX / io1 SSD (PIOPS) 256k / 64k 4,000 / 1,000 Critical high-IOPS DB; io2 = 99.999% + Multi-Attach
st1 HDD 500 500 Big sequential (logs, big data); not boot
sc1 HDD 250 250 Cold, cheapest; not boot

Instance store = ephemeral, fastest, lost on stop/terminate.

EFS vs FSx

Service Protocol/OS Use
EFS NFS / Linux Shared Linux across AZs, elastic
FSx Windows SMB / Windows Windows shares + AD
FSx Lustre Lustre HPC/ML, S3-integrated

Flashcards

Many instances, same files at once — EBS? No — EBS is single-instance/single-AZ. Use EFS (Linux) or FSx for Windows (SMB).
Default EBS volume type, and why? gp3 — baseline 3,000 IOPS/125 MiB/s, IOPS & throughput scalable independently, cheaper than gp2.
Mission-critical DB needs sustained high IOPS? Provisioned IOPS SSD (io1/io2); io2 adds 99.999% durability + Multi-Attach.
Big sequential log/big-data throughput cheaply? st1 (Throughput Optimized HDD). Note: HDDs can't be boot volumes.
Unknown/changing S3 access pattern, no retrieval fees? S3 Intelligent-Tiering.
Reproducible, infrequently accessed data, cheapest resilient-enough? S3 One Zone-IA (single AZ — fine if data can be regenerated).
Coldest, cheapest archive, hours to retrieve OK? S3 Glacier Deep Archive (min 180 days).
Shared Windows drive with Active Directory? FSx for Windows File Server (SMB).

Common traps

  • Shared files across instances → EFS/FSx, not EBS.
  • HDD (st1/sc1) can't boot; gp3 default (cheaper than gp2).
  • io2 for high IOPS DBs; st1 for sequential throughput.
  • Instance store = ephemeral (lost on stop/terminate).
  • Intelligent-Tiering for unknown patterns; IA min 30d, Glacier 90d, Deep Archive 180d.
  • One Zone-IA = single AZ (reproducible data only).