HPE7-S01 試験問題と解答 - 模擬練習 - 練習問題・模擬試験

Question No : 1
In a Retrieval Augmented Generation (RAG) pipeline, why might the system memory (RAM) requirements be significantly higher than a standard inference workload for the same model size?

A.RAG requires loading a massive vector database index into system memory to ensure low-latency retrieval of context before the LLM inference step.
B.The Embedding model used in RAG is typically 10x larger than the LL
C.RAG requires uncompressed video buffers for text generation.
D.RAG models duplicate the LLM weights into CPU RAM for redundancy.

正解:

Question No : 2
A user wants to access the Kubeflow dashboard to manage their notebook servers. They have logged into the HPE AI Essentials user interface.
Where in the navigation menu can the user find the link to launch the Kubeflow interface?

A.Data Engineering > Data Sources
B.Analytics > SQL Editor
C.Administration > Cluster Settings
D.Tools & Frameworks > Data Science

正解:

Question No : 3
An administrator needs to update the firmware on the HPE ProLiant DL325 Gen11 control nodes within an HPE Private Cloud AI cluster. They want to initiate this update from the cloud without logging into each server's iLO individually.
Which service within the HPE GreenLake unified control plane provides this server lifecycle management capability?

A.HPE GreenLake for Block Storage
B.HPE GreenLake for Compute Ops Management
C.HPE InfoSight
D.HPE AI Essentials

正解:

Question No : 4
You are planning the resource reservation for an NVIDIA NIM deployment. You check the "NVIDIA NIM Support Matrix" and identify a "Latency" profile for your specific LLM and GPU combination.
What does a "Latency" optimized profile primarily aim to minimize?

A.Time to First Token (TTFT) and Inter-Token Latency (ITL).
B.The total number of concurrent requests per second (RPS).
C.The disk space used by the model weights.
D.The total number of GPUs required.

正解:

Question No : 5
A European healthcare provider processes patient data that is subject to strict GDPR regulations. They are considering using a public cloud AI service but are concerned about "Data Sovereignty."
Why does an on-premises HPE Private Cloud AI solution offer a superior architectural fit for this specific compliance requirement compared to a standard public cloud SaaS offering?

A.It prevents the AI model from learning any medical terminology.
B.It automatically redacts all patient names before processing.
C.It uses a special "GDPR-certified" hard drive firmware that public clouds do not have.
D.It guarantees physical data location, ensuring that patient data never leaves the specific data center jurisdiction (e.g., the EU), whereas public clouds may replicate data across regions for availability.

正解:

Question No : 6
HPE Cray storage systems follow specific design mandates to solve common HPC storage pain points. One such mandate is "Medianomics."
What is the primary architectural benefit of Medianomics in HPE Cray storage systems?

A.It restricts the system to use only "Medium" endurance drives to lower costs.
B.It requires all data to be stored on tape media for economic reasons.
C.It ensures that all data is automatically compressed by 50% to save media space.
D.It enables the system to provide both fast NVMe flash (for performance) and cost-effective HDD (for capacity) pools within the same file system, optimizing the cost-to-performance ratio.

正解:

Question No : 7
You are explaining the "Zero Trust" identity model of HPE AI Essentials to a customer. They ask how workloads prove their identity to each other to establish mTLS sessions, rather than relying on simple IP allow-lists.
Which specific framework and component are responsible for minting and distributing the short-lived SVIDs (SPIFFE Verifiable Identity Documents) to the workloads?

A.Vault
B.SPIRE (SPIFFE Runtime Environment)
C.Keycloak
D.Cert-Manager

正解:

Question No : 8
An Observability Engineer is designing a custom alert rule to detect when the API server error rate exceeds a certain threshold.
Which query language must be used to define the condition expression (e.g., avg(http_requests_total{status_code="500"}) > 5) inside the rule definition?

A.SQL (Structured Query Language)
B.Lucene Query Syntax
C.PromQL (Prometheus Query Language)
D.GraphQL

正解:

Question No : 9
Why might an administrator choose to use NVIDIA Data Center GPU Manager (DCGM) instead of just nvidia-smi for cluster-wide GPU management?

A.nvidia-smi is deprecated and no longer supported on H100 GPUs.
B.nvidia-smi requires a graphical desktop environment.
C.DCGM runs on the CPU only and does not query the GP
D.DCGM provides advanced diagnostics, health checks, and policy-based management for clusters, whereas nvidia-smi is primarily a single-node query tool.

正解:

Question No : 10
An AI Administrator is monitoring the status of the KubeRay cluster running in HPE AI Essentials. They notice that while the Ray Dashboard is accessible, the "Cluster Metrics" visualizations (CPU/GPU usage graphs) are empty or missing.
What is the likely cause of this missing metrics data in the Ray Dashboard?

A.The Prometheus Node Exporter is disabled on the worker nodes.
B.Grafana has not been installed on the Ray head node.
C.The OTel collector has not been configured in the global settings.
D.The user does not have the "Cloud Administrator" role required to view metrics.

正解:

Question No : 11
You are integrating a Kubernetes cluster with HPE Alletra Storage MP. The customer requires that developers be able to dynamically provision storage, but you need to ensure different classes of service (e.g., "Gold" vs. "Bronze") map to specific Quality of Service (QoS) priority levels on the array.
Where do you define these specific volume attributes (like QoS limits or CPG selection)?

A.In the hpe-csi-driver Helm chart values.
B.In the parameters section of the StorageClass definition.
C.In the Persistent Volume Claim (PVC) metadata.
D.In the Pod specification under volumeMounts.

正解:

Question No : 12
An AI Administrator needs to check the GPU status of a specific training pod (training-job-0) that is unresponsive and does not have an interactive shell or web UI exposed.
Which kubectl command allows the administrator to remotely execute nvidia-smi inside the target pod to check for zombie processes or memory issues?

A.kubectl describe pod training-job-0 -n <namespace>
B.kubectl logs training-job-0 -n <namespace> | grep gpu
C.kubectl get gpu training-job-0 -n <namespace>
D.kubectl exec -it training-job-0 -n <namespace> -- nvidia-smi

正解:

Question No : 13
An IT Director is concerned about the ongoing operational overhead of maintaining an AI cluster. They are comparing a "Build-Your-Own" (BYO) strategy versus HPE Private Cloud AI.
Who is responsible for validating and providing full-stack software updates (firmware, OS, AI frameworks) in the HPE Private Cloud AI turnkey model compared to BYO?

A.In the turnkey model, NVIDIA manages hardware firmware, while HPE manages software.
B.In the turnkey model, the customer is responsible for validating driver compatibility before installing.
C.In both models, HPE manages all updates via GreenLake.
D.HPE provides and handles the full-stack updates for the turnkey solution; in BYO, the internal IT staff must research, procure, and install updates for every component manually.

正解:

Question No : 14
You are navigating the HPE AI Essentials user interface to review the metrics from a recent training run.
To verify the specific hyperparameters (e.g., learning_rate, batch_size) logged during the run, which section of the interface should you access?

A.Tools & Frameworks > Ray Dashboard
B.Analytics > Spark Applications
C.Data Science > Experiments > [Select Experiment] > [Select Run]
D.Data Science > Models > [Select Model]

正解:

Question No : 15
A customer observes that their legacy InfiniBand cluster experiences significant performance jitter (run-to-run variability) when multiple workloads run simultaneously. They are considering HPE Slingshot.
Which specific capability of the HPE Slingshot QoS (Quality of Service) architecture directly addresses the "noisy neighbor" problem where one workload impacts another?

A.The ability to detect congestion at the switch level and send backpressure signals specifically to the source NIC causing the congestion, throttling it before it impacts the shared fabric.
B.The use of a central subnet manager to recalculate routes every 5 minutes.
C.The use of Deep Packet Inspection (DPI) to drop low-priority packets.
D.The ability to reserve physical cables for specific jobs.

正解:

HP HPE7-S01 問題練習