A biotech company is designing a new genomics analysis platform. Thousands of parallel batch jobs must read from and write to a large shared dataset (hundreds of TB) with extremely high aggregate throughput and sub-millisecond latency during compute-intensive alignment steps. Jobs are containerized and run in bursts of up to 5,000 vCPUs, then the cluster scales to zero when idle. The dataset originates in an Amazon S3 bucket, and results must be written back to S3 for long-term retention. The team wants a fully managed solution that minimizes idle cost. Which storage and compute design best meets these requirements?