問題1
An AI infrastructure engineer is troubleshooting a poorly performing distributed training job. The job is running across multiple nodes, each equipped with powerful GPUs. The engineer observes that overall GPU utilization is unexpectedly low. System-level monitoring on the compute nodes provides the following metrics during a training run.
avg_gpu_utilization: 25%
avg_cpu_iowait_percent: 65%
avg_network_bandwidth_util: 95% (on a 10GbE network)
storage_array_latency: <1ms
Given these metrics, what is the most likely bottleneck causing the low GPU utilization?
An AI infrastructure engineer is troubleshooting a poorly performing distributed training job. The job is running across multiple nodes, each equipped with powerful GPUs. The engineer observes that overall GPU utilization is unexpectedly low. System-level monitoring on the compute nodes provides the following metrics during a training run.
avg_gpu_utilization: 25%
avg_cpu_iowait_percent: 65%
avg_network_bandwidth_util: 95% (on a 10GbE network)
storage_array_latency: <1ms
Given these metrics, what is the most likely bottleneck causing the low GPU utilization?
正確答案: B
問題2
A national research laboratory is investing in a turnkey AI infrastructure solution. Their primary goal is to eliminate the complexity and risk of designing and integrating the compute, network, and storage components themselves. The solution must be pre-validated by the vendors to deliver predictable, linear performance as they scale from one to multiple compute nodes. Which two options represent this type of pre-validated, converged infrastructure solution for AI? (Choose two)
A national research laboratory is investing in a turnkey AI infrastructure solution. Their primary goal is to eliminate the complexity and risk of designing and integrating the compute, network, and storage components themselves. The solution must be pre-validated by the vendors to deliver predictable, linear performance as they scale from one to multiple compute nodes. Which two options represent this type of pre-validated, converged infrastructure solution for AI? (Choose two)
正確答案: B,C
問題3
An organization's AI platform team needs to provide two distinct tiers of storage for their data scientists on a single Kubernetes cluster:
1. 'gold-tier': Extremely low-latency storage for active model training, using an all-flash NetApp ASA system.
2. 'bronze-tier': Cost-effective, high-capacity storage for data staging and archiving, using a NetApp StorageGRID system.
How should the platform team configure NetApp Trident to meet these requirements? (Select all that apply.)
An organization's AI platform team needs to provide two distinct tiers of storage for their data scientists on a single Kubernetes cluster:
1. 'gold-tier': Extremely low-latency storage for active model training, using an all-flash NetApp ASA system.
2. 'bronze-tier': Cost-effective, high-capacity storage for data staging and archiving, using a NetApp StorageGRID system.
How should the platform team configure NetApp Trident to meet these requirements? (Select all that apply.)
正確答案: A,B,C,F
問題4
A training job on one of the NVIDIA DGX servers is running slowly. A performance engineer runs the 'dstat' command on the server and captures the following output during the job execution.
-total-cpu-usage- -dsk/total- -net/total- paging-- system-- usr sys idl wai hiq siq| read writ| recv send| in out | int csw 15 5 70 10 0 0| 1.2G 15.0M| 1.2G 12.0M| 0 0 | 15k 35k
14 6 69 11 0 0| 1.2G 14.3M| 1.2G 11.8M| 0 0 | 14k 33k
16 5 68 11 0 0| 1.2G 16.1M| 1.2G 13.1M| 0 0 | 16k 36k
The server is connected to the NetApp ASA via a 100GbE (12.5 GB/s) network.
What is the most likely performance bottleneck based on this data?
A training job on one of the NVIDIA DGX servers is running slowly. A performance engineer runs the 'dstat' command on the server and captures the following output during the job execution.
-total-cpu-usage- -dsk/total- -net/total- paging-- system-- usr sys idl wai hiq siq| read writ| recv send| in out | int csw 15 5 70 10 0 0| 1.2G 15.0M| 1.2G 12.0M| 0 0 | 15k 35k
14 6 69 11 0 0| 1.2G 14.3M| 1.2G 11.8M| 0 0 | 14k 33k
16 5 68 11 0 0| 1.2G 16.1M| 1.2G 13.1M| 0 0 | 16k 36k
The server is connected to the NetApp ASA via a 100GbE (12.5 GB/s) network.
What is the most likely performance bottleneck based on this data?
正確答案: C
問題5
The Chief Information Security Officer (CISO) is concerned about the risk of a ransomware attack encrypting the critical vector database hosted on the NetApp AFF A-Series. The CISO wants a solution that can proactively detect and block a live attack in real-time, not just recover from a backup after the fact.
Which NetApp security feature should the architect enable on the vector database volume to meet this requirement?
The Chief Information Security Officer (CISO) is concerned about the risk of a ransomware attack encrypting the critical vector database hosted on the NetApp AFF A-Series. The CISO wants a solution that can proactively detect and block a live attack in real-time, not just recover from a backup after the fact.
Which NetApp security feature should the architect enable on the vector database volume to meet this requirement?
正確答案: D
問題6
An AI architect needs to design a complete, end-to-end data pipeline for a new generative AI application at a financial services firm. The application will allow internal analysts to query a massive, 500 TB archive of historical market data and reports to generate summaries. The firm has the following environment and requirements:
Data_Sources: A mix of on-premises ONTAP filers and StorageGRID S3 buckets.
Requirement_1: All queries must be answered using only the private data archive.
Requirement_2: All generated summaries must provide citations to the source reports.
Requirement_3: All data containing client PII must be identified and excluded from the LLM context.
Requirement_4: The solution must be cost-effective for the large, mostly-read data archive.
Which set of actions and technologies constitutes the most robust and compliant solution?
(Select all that apply.)
An AI architect needs to design a complete, end-to-end data pipeline for a new generative AI application at a financial services firm. The application will allow internal analysts to query a massive, 500 TB archive of historical market data and reports to generate summaries. The firm has the following environment and requirements:
Data_Sources: A mix of on-premises ONTAP filers and StorageGRID S3 buckets.
Requirement_1: All queries must be answered using only the private data archive.
Requirement_2: All generated summaries must provide citations to the source reports.
Requirement_3: All data containing client PII must be identified and excluded from the LLM context.
Requirement_4: The solution must be cost-effective for the large, mostly-read data archive.
Which set of actions and technologies constitutes the most robust and compliant solution?
(Select all that apply.)
正確答案: A,B,C,D
問題7
An AI research team is experiencing slow model training times. Their performance monitoring indicates that the GPUs are frequently idle, waiting for data. They want to implement a single technology change to create a more direct data path between their storage and GPUs.
Their current setup is as follows:
Compute: Server with NVIDIA A100 GPUs
Storage: NetApp AFF A-Series (All-Flash)
Network: 100GbE Ethernet
Data_Path: Storage -> Host CPU/Memory -> GPU Memory
Which technology should the architect recommend to specifically address this data path inefficiency?
An AI research team is experiencing slow model training times. Their performance monitoring indicates that the GPUs are frequently idle, waiting for data. They want to implement a single technology change to create a more direct data path between their storage and GPUs.
Their current setup is as follows:
Compute: Server with NVIDIA A100 GPUs
Storage: NetApp AFF A-Series (All-Flash)
Network: 100GbE Ethernet
Data_Path: Storage -> Host CPU/Memory -> GPU Memory
Which technology should the architect recommend to specifically address this data path inefficiency?
正確答案: D