AWS Certified Solutions Architect - Associate SAA-C03 Practice Test 71
Question 700
A machine-learning inference container requires 45 GiB of GPU-resident weights and activations per model replica, and latency rules out paging model layers to host RAM. A proposed instance has four 12-GiB GPUs, while the inference framework runs each model replica on one GPU and does not support tensor or pipeline parallelism. Which design decision resolves the memory constraint?
- Place four full model replicas on the four 12-GiB GPUs and enable EC2 enhanced networking to address the per-replica GPU allocation failure.
- Select a larger host-memory EC2 instance with the same 12-GiB GPUs and increase CPU-to-GPU PCIe transfer bandwidth to provide 45 GiB of device-resident VRAM.
- Increase the number of 12-GiB GPUs and treat aggregate VRAM as a contiguous address space, leaving the one-GPU-per-replica model unchanged.
- Select an accelerator with sufficient usable VRAM per model replica, or redesign and validate inference software for supported model parallelism.
Explanation
Correct Answer
D:The resident weights plus working set must fit in accessible accelerator memory. A larger single device meets the current software model; validated model sharding is the alternative if supported.
Incorrect Answers
A:Enhanced networking affects packet processing and network performance. It does not let a 45-GiB resident model fit in a 12-GiB device allocation when each replica is assigned one GPU.
B:Host RAM capacity and PCIe throughput do not increase the physical device VRAM available to a GPU-resident replica, and the scenario explicitly excludes paging as a latency workaround.
C:Multiple GPUs provide separate device memories. Without supported distribution across devices, an individual replica remains constrained by one 12-GiB card.
Question 701
A biotech firm runs GPU-heavy simulation jobs for five hours every Friday. Its dedicated GPU fleet remains powered on all week. The jobs can queue and checkpoint if interrupted. Which model improves utilization?
- Queue jobs through AWS Batch or suitable orchestration and launch GPU compute only during scheduled demand, optionally using interruption-tolerant capacity.
- Continue the always-on GPU fleet at maximum size because short weekly demand cannot be queued.
- Switch every job to a one-hour Lambda GPU invocation with no supported accelerator hardware.
- Buy a long-duration Compute Savings Plan sized for the continuously running GPU fleet, leaving all unused accelerator hours unchanged.
Explanation
Correct Answer
A:Elastic batch provisioning avoids paying for long idle GPU periods while preserving job retries and output persistence.
Incorrect Answers
B:The workload explicitly tolerates delayed start and does not require a persistent online fleet.
C:Lambda does not offer arbitrary dedicated GPUs for these training workloads.
D:Commitment discounts can reduce rates on eligible usage but do not remove the many idle runtime hours. Batch scheduling aligns accelerator instance-hours with actual execution instead of merely discounting waste.
Question 702
AWS Batch queues a memory-intensive image-processing job that requests only 1 GiB RAM. It starts, then repeatedly exits with an out-of-memory error. The compute environment has larger instances available. What should be corrected?
- Raise only the queue priority and keep the inadequate memory request.
- Set realistic per-job CPU and memory resource requests matching measured peak usage and compatible compute environments.
- Change the output S3 object's storage class from Standard to Glacier to add process RAM.
- Increase the number of job definitions without adjusting any resource requirements.
Explanation
Correct Answer
B:Batch schedules according to declared resources; underrequested memory can let a job start in an unsuitable allocation.
Incorrect Answers
A:Scheduling priority cannot make a memory-constrained container fit its runtime working set.
C:Output object pricing is unrelated to the job's memory limit.
D:Multiple definitions do not prevent the same out-of-memory failure.
Question 703
Three research teams submit AWS Batch jobs to one active queue and one shared compute environment. A large low-priority team fills the queue with thousands of tasks, delaying small urgent runs for other teams. The requirement is fair access over time while retaining one queue. What should be configured?
- Use only FIFO scheduling and double the retry count for every small job so the large team’s submissions stay at the front.
- Enable EC2 instance termination protection on all Batch workers, which assigns resource fairness to each submitting group.
- Attach an AWS Batch fair-share scheduling policy to the queue and submit jobs with appropriate share identifiers and weight settings.
- Put all jobs in the same dependency chain so every team must finish before any other team can schedule work.
Explanation
Correct Answer
C:Fair-share scheduling controls scheduling order and resource share among teams using share identifiers and policy factors.
Incorrect Answers
A:FIFO does not enforce proportional team fairness; retries do not resolve starvation when many earlier jobs are runnable.
B:Instance termination protection affects individual EC2 instances, not Batch job allocation among independent share identifiers.
D:A global serial dependency chain reduces parallelism and does not allocate fair shares across unrelated teams.
Question 704
A research lab already submits container jobs through AWS Batch. A fluid-dynamics solver uses MPI to coordinate 16 compute nodes throughout one tightly coupled job. Submitting sixteen independent single-node or array jobs cannot satisfy the solver’s inter-node coordination contract. Which Batch configuration meets the compute model?
- Submit an AWS Batch array job with sixteen independent children and assume arrays provide synchronized MPI ranks and tightly coupled network placement automatically.
- Switch to AWS Batch Fargate and select a 16-node multi-node parallel job definition because Fargate supplies the managed Batch MPI node model.
- Use a single-node Batch job with one larger vCPU quota and expect that it will transparently span sixteen separate EC2 hosts during execution.
- Create an AWS Batch multi-node parallel job definition on a suitable EC2 compute environment with node ranges and the required MPI communication libraries.
Explanation
Correct Answer
D:AWS Batch multi-node parallel jobs provide a main node and child nodes as one coordinated EC2-backed job. The application must still supply its MPI libraries and account for network/placement constraints.
Incorrect Answers
A:Array jobs run many independent related tasks, which suits embarrassingly parallel workloads. They do not inherently form one coordinated multi-host MPI job with shared rank discovery.
B:AWS Batch multi-node parallel jobs are not supported on Fargate compute environments. Fargate simple jobs are suitable for independent containers, not this Batch MNP requirement.
C:Increasing resources requested by a single-node job does not convert it into several networked nodes. The solver needs a job definition with explicit multi-node placement and coordination.
Question 705
A team needs to run a distributed Spark transformation across hundreds of S3 partitions. It is considering manually configuring Hadoop daemons, dependency management and worker scaling on generic EC2. Which managed service should it evaluate?
- Use Amazon EMR with supported Spark runtime and cluster/service configuration appropriate to the workload.
- Use one EC2 instance with a large EBS volume and assume it has the distributed execution properties of a Spark cluster.
- Use Amazon Route 53 to orchestrate Spark executors because DNS provides cluster scheduling.
- Use S3 Glacier Deep Archive as the active Spark shuffle medium with immediate random access.
Explanation
Correct Answer
A:EMR supplies managed distributed data processing components, reducing the need to assemble and administer the Spark cluster manually.
Incorrect Answers
B:A larger local disk does not distribute data and tasks across hundreds of worker processes.
C:DNS resolution does not provision or schedule distributed Spark data transformations.
D:An archival object tier is not a low-latency shuffle filesystem.
Question 706
An EMR Spark cluster is sized for its worst-case monthly pipeline but sits almost idle between runs. The business expects large variable daily loads and wants controlled scaling. What should change?
- Fix core and task node counts at maximum for all hours because Spark can never scale its cluster after creation.
- Use supported EMR managed scaling or elastic cluster lifecycle controls based on workload demand.
- Switch peak-sized EMR task nodes from On-Demand to Spot while leaving their desired instance count fixed during idle periods.
- Enable Spark executor dynamic allocation but retain a permanently fixed maximum EC2 instance fleet without enabling EMR cluster-level scaling.
Explanation
Correct Answer
B:Distributed worker capacity can follow job pressure instead of remaining at the expensive fixed peak.
Incorrect Answers
A:Supported EMR deployment modes allow managed or explicit capacity changes.
C:Spot pricing may lower compute unit cost but does not automatically release idle cluster capacity. The requirement concerns elasticity under variable demand, so managed scaling or lifecycle controls remain necessary.
D:Spark dynamic allocation can change active executors, but it does not guarantee the underlying EMR EC2 capacity shrinks and stops billing. The cluster needs appropriate capacity scaling or termination policies.
Question 707
A transient EMR job writes its only output to an attached worker's ephemeral HDFS files and the cluster terminates immediately after the run. Analysts cannot retrieve the result. Which data-placement practice prevents the loss?
- Keep outputs only on temporary worker disks and assume a terminated EMR cluster retains them forever.
- Shorten the EMR cluster name so its transient filesystem becomes permanent.
- Write durable inputs and final outputs to S3 before terminating transient EMR compute.
- Change only the worker security groups to allow wider SSH access after the cluster is deleted.
Explanation
Correct Answer
C:S3 decouples persisted datasets from EMR cluster lifetime; instance-local and transient HDFS storage may disappear with the cluster.
Incorrect Answers
A:Terminating transient hosts can remove the only copies of their local data.
B:Names do not control durability or post-termination data retention.
D:Network access rules cannot retrieve local bytes from terminated hosts.
Question 708
A data team runs a daily transformation of one 15-MiB CSV file from S3. The code finishes in seconds on modest resources, but the current proposal deploys a large always-on EMR/Hadoop cluster. What should be evaluated first?
- Require a 100-node EMR cluster because every CSV is a Hadoop-scale input irrespective of size.
- Deploy dedicated high-memory EC2 instances around the clock without measuring the few seconds of real demand.
- Move the 15-MiB transformation to a continuously running high-capacity AWS Batch compute environment with large GPU workers to avoid choosing a small on-demand function.
- Use an appropriately sized Lambda function or managed AWS Glue job if runtime and dependencies fit, instead of a full Hadoop cluster.
Explanation
Correct Answer
D:Small short transformations can avoid the fixed complexity and idle compute overhead of distributed cluster management.
Incorrect Answers
A:A small single file does not require distributed data processing infrastructure.
B:Permanent large compute has poor utilization for an infrequent small task.
C:Batch can run appropriate data jobs, but a permanently provisioned oversized GPU fleet adds idle expense and operational burden for a seconds-long daily CSV transform without measured accelerator needs.
Question 709
An architect creates an EBS volume in Availability Zone us-east-1a and attempts to attach it directly to an EC2 instance in us-east-1b. The API rejects the attachment despite correct IAM rights. What constraint is relevant?
- Place an EBS volume and its attached EC2 instance in compatible same-AZ locations, or migrate data through snapshots to another AZ.
- Grant administrator IAM access to bypass the zonal block-device attachment limitation.
- Increase VPC route-table entries so EBS attachment becomes a routed cross-AZ network share.
- Enable EBS Multi-Attach and expect it to attach a single block volume from us-east-1a to an instance in us-east-1b.
Explanation
Correct Answer
A:Normal EBS block attachments are zonal; cross-AZ access requires data movement or a different shared storage architecture.
Incorrect Answers
B:Permissions cannot make a volume in one AZ directly attach to an instance in another AZ.
C:Normal EBS volumes are not mounted through VPC IP routing as remote NFS storage.
D:EBS Multi-Attach is a supported same-AZ block attachment capability for eligible volume types and instances; it does not permit one EBS volume to attach across Availability Zones.
Question 710
A tightly coupled simulation needs frequent low-latency messages among EC2 worker nodes. The team has measured inter-node networking as its main bottleneck and can accept a correlated failure domain. Which placement feature merits evaluation?
- Use a spread placement group because spreading nodes across racks always minimizes round-trip latency between each pair.
- Use an appropriate EC2 cluster placement group with compatible instance types and capacity planning.
- Use a partition placement group solely to force every worker onto one rack for minimum latency.
- Disable enhanced networking on the nodes to improve the exchange rate.
Explanation
Correct Answer
B:Cluster placement groups can improve inter-instance network proximity for latency-sensitive/high-throughput computation, at the cost of placement constraints.
Incorrect Answers
A:Spread placement primarily targets hardware isolation rather than tightly clustered network performance.
C:Partition groups separate partitions for failure isolation, not a guarantee of optimal all-to-all latency.
D:Network acceleration features generally help rather than hinder compatible high-rate messaging.
