Announcing the NVIDIA NVTabular Open Beta with Multi-GPU Support and New Data Loaders

Recently, NVIDIA CEO Jensen Huang announced updates to the open beta of NVIDIA Merlin, an end-to-end framework that democratizes the development of large-scale…

NVIDIA logo
NVIDIA
•
Vinh Nguyen
•11 min read•advanced•
--
•View Original
ApacheApache ArrowDaskPythonPyTorchTensorFlow

Overview

The article announces the open beta of NVIDIA NVTabular, highlighting its new multi-GPU support and optimized data loaders for deep learning recommenders. It emphasizes the improvements in ETL processes and data loading efficiency, enabling faster training of large-scale recommender systems.

What You'll Learn

1

How to utilize NVTabular for multi-GPU ETL processes in recommender systems

2

Why optimizing data loading is critical for GPU utilization in deep learning

3

How to implement custom data loaders for TensorFlow and PyTorch using NVTabular

Prerequisites & Requirements

  • Understanding of ETL processes and deep learning frameworks
  • Familiarity with NVIDIA RAPIDS and Dask libraries(optional)

Key Questions Answered

What improvements does NVTabular offer for ETL processes in recommender systems?
NVTabular enhances ETL processes by introducing multi-GPU support and optimized data loaders, significantly reducing ETL runtimes. For instance, ETL runtimes on the Criteo Terabyte Click Log dataset were reduced from 22 minutes to 14 minutes on a single V100 GPU, demonstrating improved efficiency for large datasets.
How does NVTabular optimize data loading for deep learning frameworks?
NVTabular optimizes data loading by using an iterable data loader that allows for batch processing directly to the GPU, enhancing GPU utilization. This method contrasts with traditional item-by-item loading, leading to significant speed improvements and higher GPU utilization during training.
What new operations have been added to NVTabular in this release?
The latest NVTabular release introduces several new operations, including Column Similarity, Dropna, Filter, FillMedian, HashBucket, JoinGroupby, JoinExternal, LambdaOp, TargetEncoding, and DifferenceLag. These operators enhance the feature engineering capabilities for recommender systems.

Key Statistics & Figures

ETL runtime reduction
Reduced from 22 minutes to 14 minutes
This was achieved on the Criteo Terabyte Click Log dataset using a single V100 GPU.
Speedup comparison
95x speedup
This was observed using NVTabular multi-GPU on the DGX A100 compared to Spark on a four-node CPU cluster.

Technologies & Tools

Some links below are affiliate links. We may earn a commission if you make a purchase.

Etl Framework
Nvidia Nvtabular
Used for optimizing ETL processes in recommender systems.
Data Processing Library
Rapids Cudf
Facilitates GPU-accelerated data manipulation.
Parallel Computing Library
Dask
Enables scalable data processing across multiple GPUs.
Deep Learning Framework
Tensorflow
Used for training deep learning models with optimized data loading.
Deep Learning Framework
Pytorch
Also utilized for training models with NVTabular's data loaders.

Key Actionable Insights

1
Leverage NVTabular's multi-GPU support to accelerate your ETL processes for large datasets.
This is particularly beneficial when working with terabyte-scale datasets, as it can drastically reduce processing times and improve overall workflow efficiency.
2
Utilize the new data loaders in NVTabular to enhance the performance of your TensorFlow and PyTorch models.
By implementing these optimized data loaders, you can ensure better GPU utilization and faster training times, which is crucial for developing efficient recommender systems.

Common Pitfalls

1
Failing to optimize data loading can lead to underutilization of GPU resources.
This often occurs when traditional item-by-item loading methods are used, which can significantly slow down training times. Adopting NVTabular's iterable data loader approach can mitigate this issue.

Related Concepts

Etl Processes In Machine Learning
GPU Acceleration In Deep Learning
Feature Engineering Techniques