SAA-C03 ยท Elastic compute performance
Placement groups, ENA and EFA
Cluster, spread and partition placement groups, Elastic Network Adapter and Elastic Fabric Adapter for low-latency, high-throughput compute.
Key points
- Cluster: packed in one AZ, lowest latency, highest per-flow throughput (tightly coupled HPC); one request, one instance type.
- Partition: partitions never share racks, up to 7 per AZ, can span AZs; HDFS, HBase, Cassandra.
- Spread: each instance on distinct hardware, max 7 running per AZ.
- ENA: enhanced, high-bandwidth IP networking.
- EFA: OS-bypass for MPI and NCCL; that traffic cannot cross AZs.
Source: https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/placement-strategies.html
Practice questions
Which placement strategy gives tightly coupled HPC nodes the lowest network latency between each other?
- Spread
- Partition
- Cluster
- Precision time
Show answer
Answer: C. Cluster
A cluster placement group places instances in the same high-bisection bandwidth segment of one Availability Zone, giving the lowest latency. Spread and partition strategies separate instances across hardware for resilience, which works against latency. Precision time is about clock accuracy, not latency.
Which statements about EC2 placement groups are correct?
- A cluster placement group stays within one Availability Zone
- A spread placement group runs at most seven instances per AZ
- A partition placement group holds at most seven instances in total
- Mixing many instance types is recommended in cluster groups
- A spread placement group must stay in a single AZ
Show answer
Answer: A. A cluster placement group stays within one Availability Zone, B. A spread placement group runs at most seven instances per AZ
A cluster placement group cannot span Availability Zones, and a rack-level spread group supports at most seven running instances per AZ. Partition groups are limited to seven partitions per AZ, not seven instances in total, and AWS recommends one instance type in a cluster group because mixing types raises the chance of capacity errors.
Traffic that uses the OS-bypass path of an Elastic Fabric Adapter can flow between instances in different Availability Zones.
Show answer
Answer: False
False. EFA traffic cannot cross Availability Zones or VPCs and is not routable. Only the normal IP traffic from the ENA part of an EFA interface is routable, so tightly coupled EFA jobs belong in one AZ, typically in a cluster placement group.
Practice 8 questions on this topic in Bitzu
Short lessons, instant explanations and spaced review of what you get wrong. Free on iOS and Android.
