EC2 Placement Groups

A cluster placement group packs instances close together inside a single Availability Zone, ideally on the same high-bisection-bandwidth network segment. This delivers the lowest network latency and the highest packet-per-second throughput between instances, which tightly coupled workloads need. Instances in a cluster group can use Elastic Fabric Adapter (EFA), a network interface that accelerates inter-node communication for High Performance Computing (HPC). The trade-off is reduced fault isolation: because instances sit on shared, nearby hardware, a hardware failure can affect many of them at once. A cluster group cannot span Availability Zones.

A spread placement group places each instance on distinct underlying hardware, meaning each instance sits on its own rack with its own network and power source. This minimizes the chance that a single hardware failure takes down more than one instance, which suits a small number of critical instances that must stay independent. A spread group can span multiple Availability Zones within a Region, but it is limited to a maximum of seven running instances per Availability Zone per group. The strategy optimizes for availability rather than for network performance.

A partition placement group divides instances into logical segments called partitions, where each partition maps to a distinct set of racks that do not share underlying hardware (power and network) with other partitions in the same group. A failure in one partition affects only that partition. This combines fault isolation with the ability to scale to many instances, which is why large distributed and replicated workloads use it. You can have up to seven partitions per Availability Zone, and a partition group can span multiple Availability Zones in a Region. EC2 exposes the partition number to the instance, so distributed applications can place data replicas in different partitions on purpose.

Cluster fits tightly coupled, node-to-node workloads such as HPC simulations, scientific modeling, and big data jobs that exchange large volumes of traffic and benefit from the lowest latency. Spread fits a small number of instances that must remain isolated from one another, such as a set of critical application servers or replicated control-plane nodes where losing more than one at the same time is unacceptable. Partition fits large distributed and replicated systems such as HDFS (Hadoop Distributed File System), HBase, Cassandra, and Kafka, where you spread replicas across partitions so a single rack failure cannot take down all copies of the data.

Each strategy has boundaries the exam tests. A cluster group lives in one Availability Zone, and launching many instances at once is recommended because adding instances later can fail with an InsufficientInstanceCapacity error if the close-proximity hardware is full. A spread group allows a maximum of seven running instances per Availability Zone per group. A partition group supports up to seven partitions per Availability Zone. Across all types, you cannot merge placement groups, and you generally should keep instance types consistent in a cluster group for best performance. You can move or add an existing instance into a placement group, but the instance must be in the stopped state when you do so.