Amazon Instance Store is the correct choice because it provides the highest local IOPS at the lowest cost when each EC2 node already holds a full replica of the dataset.

Amazon Instance Store uses NVMe attached local storage to avoid network overhead and to deliver very low latency and high random read and write IOPS. The storage is included in the instance price which makes it cost effective for per node replicated HPC workloads and it can meet the 150,000 random IOPS requirement when you select instances with adequate instance store capacity.

Amazon EBS io2 Block Express can deliver extremely high IOPS but it is network attached and requires provisioning and billing for IOPS and capacity which usually makes it more expensive for per node replicas. EBS does provide durable volumes that persist across instance stops so choose it when persistence or reattachment is required.

Amazon S3 with byte-range fetch is object storage and it does not provide block level semantics or predictable per volume IOPS which makes it unsuitable for a low latency, high IOPS block workload.

Amazon FSx for Lustre is a high throughput shared filesystem designed for HPC but it introduces extra network latency and cost which are unnecessary when every node already maintains a full replica and requires maximum local IOPS.

Full AWS Practitioner Certification Question