An architect is designing a scalable, automated MLOps platform using Kubeflow on a Kubernetes cluster. The platform must support the entire AI lifecycle for multiple teams, with different storage requirements at each stage.
The key requirements are:
- Data Ingestion: A pipeline step needs a shared, read-write volume accessible by multiple pods to stage raw data.
- Experimentation: Data scientists need individual, isolated volumes for their Jupyter notebooks.
- Training: Distributed training jobs require a high-performance, parallel-access filesystem for reading training data.
- Automation: All storage must be provisioned automatically via Kubeflow pipeline definitions without manual intervention.
Which combination of technologies and configurations would create the most effective solution?

• A. Create a single, large NFS volume and mount it to all pods using a static PersistentVolume.
• B. Use the NetApp DataOps Toolkit for all storage provisioning, bypassing Trident and Kubernetes PVCs.
• C. Rely on hostPath volumes for all storage to ensure the highest performance.
• D. Configure multiple Trident backends (e.g., 'ontap-nas' for standard volumes, 'ontap-nas-flexgroup' for parallel access) and corresponding StorageClasses.
• E. Use the NetApp DataOps Toolkit for Python within the Kubeflow pipeline components to dynamically create and manage Trident PVCs for each stage.

Which AI technology is used to generate new, never-before-seen content such as images or text?

• A. Supervised AI
• B. Reinforcement AI
• C. Predictive AI
• D. Generative AI

An AI training job is running slower than expected. The infrastructure team suspects a compute bottleneck. The job involves processing high-resolution images with a complex convolutional neural network (CNN). They review the logs from the training script.
Epoch 1/100 - 3600s - loss: 1.253 - acc: 0.54
...
CPU_Utilization_During_Epoch: 99% (all cores)
GPU_0_Utilization_During_Epoch: 8%
GPU_1_Utilization_During_Epoch: 7%
...
Epoch 2/100 - 3610s - loss: 1.102 - acc: 0.61
What is the most likely cause of the poor training performance?

• A. The model is too simple and does not effectively utilize the GPU's parallel processing capabilities.
• B. The network is saturated, preventing the GPUs from receiving data.
• C. The storage system cannot deliver data fast enough to the compute node.
• D. The training script is CPU-bound, likely performing data augmentation or preprocessing on the CPU instead of offloading it to the GPUs.

An AI platform is suffering from poor performance during distributed training jobs. The training data resides on a single, large NFS volume. Monitoring shows that while the overall network throughput to the storage system is high, individual GPU nodes experience significant I/O wait times, and the single ONTAP volume is becoming a performance bottleneck. The goal is to re- architect the storage layout to maximize read parallelism and throughput for the training cluster.
Which two actions should the architect take to address this performance bottleneck? (Choose 2.)

• A. Implement a NetApp FlexGroup volume to spread the dataset across multiple constituent volumes and aggregates.
• B. Replace the NFS protocol with iSCSI for all training data access.
• C. Use NetApp FlexCache to create a local cache of the training data on each compute node.
• D. Increase the number of network ports connected to the storage controller.
• E. Enable QoS maximums on the training volume to limit its IOPS.