Accelerating Batch Data Processing on GPUs with Rapids Accelerator for Apache Spark

Overview

• Challenges of big data processing:
  • Exponential growth of data from various sources (internet, consumer devices, IoT)
  • Need for high-quality data at low latency
• Approaches to handle data growth:
  • Horizontal scaling of compute resources
  • Downsampling data
  • Longer processing times
• Proposed solution: Accelerate batch data processing using Nvidia GPUs and the Rapids Accelerator for Apache Spark

Apache Spark 3 and Rapids Accelerator

• Apache Spark 3 features that enable GPU processing:
  • Resource-aware scheduling
  • Columnar data processing
  • Plugin architecture
• Rapids Accelerator for Apache Spark:
  • Plug-in that works with different Apache Spark distributions (EMR, Databricks, etc.)
  • Seamless integration with existing user workflows (SQL, DataFrames)
  • Transparent GPU acceleration, with fallback to CPU for unsupported operations

Performance and Cost Savings

• Benchmarks:
  • NDS (Nvidia Decision Support Benchmark) on AWS EC2 instances:
    • Up to 9x faster performance on GPU, with 4.8x average speedup and 72% cost savings
  • NDS on AWS EMR:
    • Up to 3.5x faster performance on GPU, with 1.3x average speedup and 14% cost savings
• Transaction Fraud Use Case:
  • 14x speedup and 87% cost savings when running on GPU cluster

Identifying Workloads for GPU Acceleration

• Spark Rapids User Tools:
  • Open-source tool to analyze Spark event logs and recommend workloads for GPU acceleration
  • Provides recommendations on cluster configuration and job tuning

Getting Started

• Documentation:
  • Nvidia Rapids Accelerator for Apache Spark documentation
  • AWS EMR documentation on using the Rapids Accelerator
• Spark Rapids User Tools:
  • Available on Python Package Installation site
  • Notebook for running the tool on EMR

Conclusion

• Exponential data growth can be addressed by accelerating batch data processing on GPUs using the Rapids Accelerator for Apache Spark
• Significant performance improvements and cost savings can be achieved by identifying the right workloads for GPU acceleration
• The open-source Spark Rapids User Tools can help in this process by analyzing Spark event logs and providing recommendations