Examkin

Question

Difficulty: Very hardHigh-Performing and Elastic Compute Solutions

A financial modeling firm is designing a high-performance compute (HPC) environment on AWS to run tightly-coupled risk simulations. The workload uses Message Passing Interface (MPI) and requires node-to-node network latency of less than 1010 microseconds. The cluster must dynamically scale between 2020 and 8080 compute-optimized EC2 instances based on queue depth. During load testing, the team observes that scaling out the Auto Scaling group incrementally results in frequent launch failures with `InsufficientInstanceCapacity` errors, and the network performance between newly launched instances does not meet the low-latency SLA.

Which architecture should a Solutions Architect recommend to resolve the capacity errors and maintain the required low-latency performance?

  1. Deploy the EC2 instances in a single Cluster Placement Group within a single Availability Zone. Use a Launch Template that specifies an Elastic Fabric Adapter (EFA) enabled instance type, and associate the Auto Scaling group with an On-Demand Capacity Reservation targeted for the placement group to guarantee that the maximum capacity of 8080 instances is pre-allocated.Answer
  2. B
    Deploy the EC2 instances across multiple Availability Zones using a Spread Placement Group. Enable Elastic Fabric Adapter (EFA) in the Launch Template, and configure a scaling policy based on the average network in/out metric to distribute the simulation nodes evenly.
  3. C
    Deploy the EC2 instances in a Partition Placement Group with 77 partitions within a single Availability Zone. Configure an Application Load Balancer to distribute the simulation control traffic, setting the target group health check to the default port 8080 while the simulation application runs on port 80808080.
  4. D
    Deploy the EC2 instances in a Cluster Placement Group spanning 33 Availability Zones. Use a Launch Template that enables Elastic Network Adapter (ENA) Express to automatically optimize node-to-node communication latency and throughput across the zones.

Answer

Deploy the EC2 instances in a single Cluster Placement Group within a single Availability Zone, using a Launch Template that specifies an Elastic Fabric Adapter (EFA) enabled instance type, and associate the Auto Scaling group with an On-Demand Capacity Reservation targeted for the placement group.
Tightly-coupled HPC workloads using MPI require ultra-low latency, which is achieved by deploying EC2 instances in a Cluster Placement Group in a single Availability Zone and using Elastic Fabric Adapter (EFA). Launching instances incrementally into a cluster placement group often leads to capacity fragmentation and `InsufficientInstanceCapacity` errors if the underlying physical rack lacks contiguous capacity. Creating an On-Demand Capacity Reservation (ODCR) targeted for the cluster placement group ensures that the required slot of 8080 instances is pre-allocated, resolving the launch failures during scale-out events.

Step-by-Step Solution

1
Identify latency and communication requirements
Tightly-coupled HPC workloads using MPI require node-to-node latency of less than 1010 microseconds, which demands a Cluster Placement Group in a single Availability Zone combined with Elastic Fabric Adapter (EFA) for OS-bypass.
This establishes the baseline compute and network topology optimized for tightly-coupled workloads.
2
Analyze the cause of scale-out capacity failures
Launching instances incrementally in a Cluster Placement Group leads to capacity fragmentation. The physical rack containing the placement group may run out of contiguous slots, causing `InsufficientInstanceCapacity` errors.
This explains why standard Auto Scaling scale-out fails even if general regional capacity exists.
3
Select the correct mitigation strategy
By using an On-Demand Capacity Reservation (ODCR) targeted for the specific Cluster Placement Group, AWS reserves the maximum 8080 instance slots upfront, ensuring subsequent scale-out requests succeed without capacity errors.
This guarantees availability of the required compute slots inside the strict physical boundaries of the placement group.

Key Concept

Optimizing compute environments for tightly-coupled HPC workloads using Cluster Placement Groups, Elastic Fabric Adapter (EFA), and On-Demand Capacity Reservations.
Estimated Time:3m 0s
Rate this question

Topics

High-Performing and Elastic Compute Solutions

More questions from this topic