Kvíz spojený s průvodcem · Střední úroveň

Kvíz o opravě úzkých míst načítání dat

Diagnostikujte zpoždění vstupního kanálu a vylaďte počet pracovníků, předběžné načítání, připnutou paměť, rozložení úložiště a přenosy pomocí naměřených důkazů.

Otázka 1 z 8

Which symptom suggests the input pipeline is limiting training?

Náhled každé otázky8 otázky

Odpovědi a vysvětlení se zobrazují, jak budete v kvízu.

  1. Which symptom suggests the input pipeline is limiting training?
    • GPU compute stays continuously active while the next batch is prepared
    • The loader places each batch in the device queue before current computation ends
    • The GPU waits between steps while batches are prepared
    • A full prefetch queue is available while the GPU remains idle
  2. Why can increasing DataLoader workers eventually reduce performance?
    • More workers always increase throughput without resource cost
    • Workers can contend for CPU, memory, or storage resources
    • Worker count changes model weights
    • Workers guarantee every data source becomes sequential
  3. What does pinned host memory help with in a compatible workflow?
    • Reduce file-read latency from remote object storage
    • Lower CPU image or audio decoding time
    • Increase the GPU’s physical VRAM capacity
    • Improve host-to-GPU data transfer in a compatible workflow
  4. How does prefetching affect batch preparation?
    • Prepare future batches before the model needs them
    • Keep the entire dataset resident in GPU memory
    • Skip preprocessing for all later epochs
    • Start reading each next batch only after the current device step ends
  5. Why might a dataset of many tiny files load slowly?
    • When every read is a large sequential transfer that saturates bandwidth
    • When model kernels, rather than storage access, saturate GPU compute
    • When metadata lookups and file opens dominate reads
    • When optimizer state occupies accelerator memory
  6. What should be verified after adding a cache or faster transform?
    • Whether the worker count matches the previous run
    • Confirm data values and labels still match the intended pipeline
    • Check that cache capacity uses all available RAM
    • Whether GPU utilization rises after any transformation
  7. When can non-blocking device copies improve a pipeline?
    • Whenever any tensor is copied, whether or not execution overlaps
    • Only when model weights are frozen
    • When the host has no memory pressure
    • When transfer can overlap with compute under the runtime and stream setup
  8. What should be done before changing several loader settings at once?
    • Profile a baseline and vary settings in controlled comparisons
    • Increase worker and prefetch settings together without measuring
    • Optimize kernels before measuring input wait time
    • Enable pinned memory for every device unconditionally