Compute for High Performance
EC2 Instance Families:
- T — burstable (t3, t4g). Dev/test.
- M — balanced (m5, m6i). General production.
- C — compute-optimized (c5, c6i). High CPU: batch, gaming, ML inference.
- R — memory-optimized (r5, r6i). In-memory DBs, SAP.
- I — storage-optimized (i3, i4i). High random IOPS: NoSQL, Elasticsearch.
- P/G — GPU. ML training, graphics.
EC2 Placement Groups:
- Cluster — same rack, same AZ. Lowest latency, highest risk. For HPC, tightly coupled workloads.
- Spread — different hardware, different AZs. Max 7 instances per AZ per group. For critical applications requiring HA.
- Partition — spread across partitions (isolated rack sets). Up to 7 partitions per AZ. For distributed systems (Hadoop, Kafka, Cassandra).
Enhanced Networking: Elastic Network Adapter (ENA) — up to 100 Gbps. Elastic Fabric Adapter (EFA) — improved ENA for HPC (Linux only). Bypasses OS for low-latency, reliable transport. MPI support.
EC2 Purchasing Options:
- On-Demand — pay by hour/second. No commitment. Most expensive.
- Reserved Instances (RI) — 1 or 3 year. Up to 72% off. Standard (fixed) vs Convertible (can change type) vs Scheduled.
- Savings Plans — commit to $/hour. More flexible than RIs. Covers EC2, Lambda, Fargate.
- Spot — up to 90% off. Can be interrupted with 2-minute warning. For fault-tolerant, stateless workloads.
- Dedicated Hosts — physical server for you. For license compliance.
- Dedicated Instances — instance on dedicated hardware, but not reserved hardware.