DEV Community

Wakeup Flower
Wakeup Flower

Posted on

EC2 Instance Store & EBS explained

#aws
  • You launch an EC2 instance (e.g., m5d.large).
  • That instance comes with instance store included (in this case, 1 × 75 GB NVMe SSD).

You don’t create “an instance store” separately — it’s a feature of the EC2 instance you pick.

✅ Example

  • m5.large → ❌ has no instance store (only EBS).
  • m5d.large → ✅ has instance store (75 GB NVMe SSD, ephemeral).

The “d” in the name is the giveaway → d = instance store included.

🏗️ How it looks:

  1. You launch an EC2 instance type that has “d” in the family (like m5d, c5d, i3, etc.).
  2. AWS gives you:
  • The EC2 compute resources (vCPU, RAM).
  • One or more instance store volumes (NVMe or SSD, ephemeral, super fast).
  • Optionally, you can also attach EBS volumes (persistent).

Perfect — here’s a text-based comparison table so you can see all options side by side:

📊 AWS Storage Options Comparison

Storage Option Performance (IOPS/Throughput) Latency Persistence Cost Model Best For Why Not Here
EC2 Instance Store ✅ Very high IOPS (NVMe/SSD, physically attached) Ultra-low (local disks) ❌ Lost when instance stops/terminates ✅ Included in EC2 cost (no extra charge) Temporary scratch space, caches, buffers Exactly what we need: fast, temporary, cost-free
EBS gp2/gp3 (General Purpose SSD) gp2: 3 IOPS/GB (burst to 16K)
gp3: baseline 3K, up to 16K		 Single-digit ms ✅ Persists ❌ Extra cost (per GB + IOPS for gp3) General workloads (boot, dev/test, light DBs) Adds cost, not as fast as Instance Store
EBS io1/io2 (Provisioned IOPS SSD) Up to 64K IOPS (Nitro-based instances) Consistent, predictable ✅ Persists ❌ Most expensive (storage + provisioned IOPS) Mission-critical DBs, OLTP Overkill, persistence not required
EBS st1 (Throughput-Optimized HDD) Optimized for throughput (MB/s), not IOPS Higher than SSD ✅ Persists ❌ Extra cost (per GB) Big data, data warehouses, sequential reads Poor random IOPS, not fit for scratch space

1. Amazon EBS General Purpose SSD (gp2 / gp3)

  • Type: Network-attached block storage (persistent).

  • Performance:

    • gp2 → 3 IOPS per GiB baseline, up to 16,000 IOPS (burstable).
    • gp3 → Baseline 3,000 IOPS, configurable up to 16,000 IOPS.

  • Persistence: ✅ Data persists beyond instance lifecycle.

  • Cost: Additional cost (charged per GB provisioned + IOPS if gp3).

  • Best For: General workloads (boot volumes, dev/test, moderate DBs).

  • Why Not Here? Persistent storage adds unnecessary cost; performance not as high as instance store.

2. Amazon EBS Provisioned IOPS SSD (io1 / io2)

  • Type: Network-attached block storage (persistent).

  • Performance:

    • Up to 64,000 IOPS (with Nitro-based EC2).
    • Consistent, predictable IOPS delivery.

  • Persistence: ✅ Data persists.

  • Cost: Most expensive EBS option (charged for storage and provisioned IOPS).

  • Best For: Critical databases, latency-sensitive OLTP apps.

  • Why Not Here? Overkill (expensive persistence when only temporary scratch space is needed).

3. Amazon EBS Throughput Optimized HDD (st1)

  • Type: HDD-based network storage (persistent).
  • Performance: Optimized for throughput (MB/s), not IOPS.
  • Persistence: ✅ Data persists.
  • Cost: Cheaper than SSD EBS, but still extra cost vs. instance store.
  • Best For: Big Data, log processing, sequential throughput workloads.
  • Why Not Here? Poor random IOPS performance (not suitable for file scratch space).

🏆 Recommendation

For computer vision researchers doing high IOPS, temporary file processing, the best choice is Amazon EC2 Instance Store:

  • ✅ Highest performance (IOPS, latency).
  • ✅ Cost-optimal (included in instance price).
  • ✅ Fits temporary nature of workload.

Top comments (0)