Courseiva
hardMultiple ChoiceObjective-mapped

Boosting GPU Utilization with Data Format and Distributed Training

A financial services company is developing a real-time fraud detection model using XGBoost on SageMaker. They have millions of transactions daily and train a model weekly on 6 months of historical data. The training dataset is 500 GB in CSV format stored in S3. The training job uses an ml.p3.16xlarge instance with 8 GPUs, but training takes over 12 hours, which is too long for the weekly cadence. The data scientist notices that GPU utilization averages only 15% during training. The training script uses the SageMaker XGBoost container with default hyperparameters. Which combination of actions would MOST likely reduce training time? (Choose the best answer.)

Quick Answer

A GPU utilization of only 15% is the key diagnostic detail here: it tells you the bottleneck isn't compute, since the GPUs are mostly idle, it's how slowly data is getting to them, and every part of the correct answer targets that data pipeline rather than the model or algorithm. Converting the 500GB CSV dataset to Parquet reduces the amount of data that has to be read and parsed, since Parquet is a compressed, columnar format built for efficient scanning. Switching to Pipe input mode changes how SageMaker delivers that data to the training container, streaming it directly instead of first downloading the full dataset to local disk, which removes a slow, sequential download step before training can even begin. Increasing the instance count enables distributed training, spreading the workload across more GPUs so each one processes a smaller share of the data in parallel rather than one set of 8 GPUs sitting mostly idle while I/O catches up. None of these changes touch the model itself, which fits the fact that the team already used default hyperparameters and the underlying issue was never the algorithm. When a question shows low GPU utilization alongside long training times on a large on-disk dataset, look for answers that fix data format, transfer mode, and parallelism rather than the model.

⚠ Common exam trap

The trap here is that candidates focus on GPU hardware upgrades (Option A) or hyperparameter tuning (Option B) without recognizing that the root cause is data I/O inefficiency from CSV format and single-instance training, which is a classic SageMaker optimization scenario.

Answer choices

Why each option matters

Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.

Correct answer & explanation

Convert the training data to Parquet format, use Pipe input mode in the training job, and increase the instance count to run distributed training.

Converting CSV to Parquet reduces data size and improves I/O efficiency, Pipe input mode streams data directly to the algorithm without downloading, and increasing instance count enables distributed training across multiple GPUs. These changes directly address the low GPU utilization (15%) by reducing data loading bottlenecks and parallelizing computation, which is the core issue with the current single-instance, CSV-based training.

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

  • Increase the instance type to ml.p3dn.24xlarge and use EFA networking.

    Why it's wrong here

    This improves networking but does not address the I/O bottleneck from CSV format and default data loading.

  • Tune hyperparameters using SageMaker Automatic Model Tuning to reduce training epochs.

    Why it's wrong here

    Hyperparameter tuning may improve convergence but not necessarily address low GPU utilization due to I/O.

  • Use SageMaker Debugger to profile the training and adjust the batch size to maximize GPU memory usage.

    Why it's wrong here

    Debugger helps identify bottlenecks but alone does not change the underlying I/O inefficiency; adjusting batch size may not be enough.

  • Convert the training data to Parquet format, use Pipe input mode in the training job, and increase the instance count to run distributed training.

    Why this is correct

    Parquet reduces data size and improves I/O; Pipe mode streams data efficiently; distributed training scales out to reduce time.

Quick reference

AWS S3 Storage Class Comparison

Storage ClassMin DurationRetrievalUse Case
S3 StandardNoneImmediateFrequently accessed data
S3 Standard-IA30 daysImmediateInfrequent access, rapid retrieval
S3 One Zone-IA30 daysImmediateNon-critical infrequent data
S3 Intelligent-TieringNoneImmediate–hoursUnknown or changing access patterns
S3 Glacier Instant90 daysMillisecondsArchive with instant retrieval
S3 Glacier Flexible90 daysMinutes–hoursArchive, flexible retrieval
S3 Glacier Deep Archive180 daysHoursLong-term compliance archive

About these practice questions

This MLA-C01 question is part of Courseiva's 835-question bank — original exam-style content with full explanations and wrong-answer analysis, never real exam questions or exam dumps. Learn why practice questions differ from exam dumps →

How Courseiva writes practice questions · Editorial policy

Same concept, more angles

1 more way this is tested on MLA-C01

These questions test the same concept from different angles. Work through them to make sure you can recognise it however the exam phrases it.

Variation 1. A company uses SageMaker to train a model with a large dataset stored in S3. They notice that the training job is taking longer than expected and the GPU utilization is low. Which action would most likely improve GPU utilization?

hard
  • A.Increase the batch size
  • B.Disable distributed training
  • C.Use a smaller instance type
  • D.Decrease the batch size

Why A: Low GPU utilization during training often indicates that the GPU is waiting for data to process, a condition known as data bottleneck. Increasing the batch size allows each training step to process more samples, which increases the computational load per step and keeps the GPU busy for longer periods, thereby improving utilization. This is especially effective when using SageMaker's managed training with large datasets stored in S3, as larger batches reduce the frequency of data loading operations.

JA

Written by Johnson Ajibi, MSc IT Security

Senior Network & Security Engineer · founder of Courseiva

This MLA-C01 practice question is part of Courseiva's free Amazon Web Services certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the MLA-C01 exam.