Guide-linked quiz · Medium Level

Storage for AI Training Clusters Quiz

Match object storage, parallel filesystems and local NVMe to AI training needs while measuring bandwidth, metadata, caching and durability.

Question 1 of 8

Why can many small files slow distributed training data access?

Preview every question8 questions

Answers and explanations appear as you take the quiz.

  1. Why can many small files slow distributed training data access?
    • When all reads are large sequential transfers that saturate bandwidth
    • When model kernels already saturate GPU compute
    • Metadata and per-file request overhead can accumulate
    • When optimizer states occupy accelerator memory
  2. Which storage type is usually designed for durable access to large collections of objects through APIs?
    • A parallel filesystem that exposes files through a POSIX-like interface
    • Object storage
    • A shared NFS volume mounted by each node
    • A host page cache
  3. How is local NVMe commonly used in a training cluster?
    • The primary durable shared repository for every node
    • A durable object store with native POSIX semantics
    • A replacement for networked storage across all compute nodes
    • Fast temporary cache or staging close to compute
  4. Why measure aggregate demand across all GPUs?
    • The cluster’s readers can collectively exceed storage throughput
    • The per-node read rate measured on an otherwise idle host
    • Average dataset size without considering reader concurrency
    • Peak disk capacity without measuring read demand
  5. When can repacking a dataset into larger shards help?
    • When data transfer is limited by a sequential-read bandwidth ceiling
    • When GPU compute is saturated while the input queue stays full
    • When small-file metadata and request overhead limits reads
    • When accelerator memory is consumed by optimizer state rather than input reads
  6. What tradeoff does checkpoint frequency involve?
    • Checkpoint data volume versus the number of GPUs
    • Recovery time versus storage traffic and pauses
    • Training-label quality versus read latency
    • GPU compute peak versus dataset licensing
  7. Why distinguish filesystem throughput from metadata performance?
    • Byte bandwidth measures the rate of file lookups and directory operations
    • Metadata performance and bulk byte throughput are the same metric
    • Small-file workloads stress only network bandwidth
    • Large sequential reads and many small-file operations stress different paths
  8. What should be checked before relying on node-local cached data after a restart?
    • Which dataset version is present and whether durable copies exist
    • Local caches always survive a worker replacement
    • The model version guarantees that cached inputs are current
    • A cache identifier proves that a backup was completed

Keep testing yourself

More quizzes picked for your level.

View all quizzes