Accelerating Apache Spark with Gluten + Velox: Vectorized Execution for Big Data at Scale

Executive Summary

As organizations process ever-growing data volumes, the demand for faster insights is relentless. Apache Spark has become the de facto standard for distributed analytics, but its row-based execution model often becomes a performance bottleneck at scale.

Acceldata’s ODP Spark with Gluten + Velox delivers a breakthrough by combining:

• Gluten
• Velox

Key performance gains on TPC-DS 100 GB benchmarks:

• ✅ 1–3× faster query execution
• ✅ 20–30 % fewer CPU cycles per row
• ✅ 15–20 % lower memory allocation pressure
• ✅ Reduced shuffle and I/O overhead

‍

Quantified ROI Example (1000-core Spark cluster)

And the best part? These improvements require no changes to your existing Spark applications.

The Problem: Spark’s Row-Based Execution

Traditional Spark processes data one row at a time inside the JVM. While flexible, this model introduces:

• High function call overhead
• JVM GC pressure from frequent object creation
• Poor CPU cache utilization
• Limited opportunities for SIMD/vectorization

This results in suboptimal performance for analytical workloads such as aggregations, joins, and window functions on large datasets.

The Solution: Vectorized Execution with Gluten + Velox

Vectorized execution processes data in columnar batches rather than rows, unlocking:

• SIMD (Single Instruction, Multiple Data) acceleration
• Improved CPU cache locality
• Reduced function call overhead
• Native (C++) execution performance

‍

Gluten Framework

Gluten acts as a bridge between Spark and native engines, replacing parts of Spark’s physical plan while maintaining full API compatibility.

Highlights:

• Catalyst integration for rule-based plan transformation
• Extensible backend support (Velox, ClickHouse, Arrow)
• Apache Arrow-based zero-copy columnar data exchange

‍

Velox Execution Engine

Velox is the native vectorized runtime originally developed by Meta for Presto. It provides:

• Columnar batch processing (1024–4096 rows per batch)
• SIMD-enabled expression evaluation and subexpression elimination
• NUMA-aware memory management
• Vectorized operators for joins, aggregates, and window functions

‍

Benchmark Results: TPC-DS on 100 GB

‍

Resource Utilization Gains:

• CPU: 20–30 % fewer cycles per row; better instruction cache hit rate
• Memory: 15–20 % lower allocation pressure; reduced JVM GC
• I/O: better predicate pushdown, improved compression, lower shuffle volume

Implementation Guide

Cluster & Storage Requirements

• Native Velox libraries deployed on all executors
• glibc ≥ 2.28, libstdc++ ≥ 11.2
• Parquet/ORC columnar storage formats
• Modern CPUs with SIMD support

Key Spark Configuration

# Enable Gluten
spark.gluten.enabled=true
spark.shuffle.manager=org.apache.spark.shuffle.columnar.ColumnarShuffleManager
# Memory tuning
spark.gluten.memory.overAcquiredMemoryRatio=0.3
spark.gluten.sql.columnar.backend.velox.memCacheSize=4g
# Vector batch tuning
spark.gluten.sql.columnar.batchSize=2048
spark.gluten.sql.columnar.backend.velox.vectorBatchSize=1024
spark.gluten.sql.columnar.backend.velox.adaptiveBatchSize=true
# NUMA optimization
spark.gluten.sql.columnar.backend.velox.numaAware=true
spark.gluten.sql.columnar.backend.velox.numaMemoryPolicy=preferred
spark.gluten.sql.columnar.backend.velox.numaLocalThreadPools=true

These settings are ideal for TPC-DS style OLAP workloads on multi-socket servers.

Real-World ROI Example

Tuning Velox for Peak Performance

• Vector batch size tuned to CPU L1 cache for better locality
• Memory pools adjusted for OLAP vs streaming workloads
• NUMA awareness enabled to reduce cross-socket memory access
• Built-in monitoring scripts for perf counters, memory usage, and Spark executor stats

Deployment & Compatibility

• No application code changes needed — unsupported queries fall back to JVM.
• Best performance on Parquet; ORC partially supported; JSON/CSV may fall back.
• Structured Streaming not yet supported.

Monitoring tips:

• Spark UI → SQL Plan Tab: look for operators like VeloxHashAggregateExec, VeloxProjectExec
• Stage & task metrics show distinctive CPU and memory patterns for native execution

Roadmap

Gluten + Velox is evolving rapidly:

• Broader SQL function coverage
• Streaming support under development
• GPU acceleration integration
• Advanced compression & query compilation optimizations
• Deep Iceberg, Delta Lake, and Hudi support

Conclusion

Acceldata ODP Spark with Gluten Velox marks a new era of vectorized, native execution for Spark. Enterprises can expect:

• 1–3× performance improvements on analytical queries
• Significant infrastructure savings
• Better CPU, memory, and I/O efficiency
• No application code rewrites

Gluten/Velox is production-ready today for OLAP workloads and is rapidly expanding its capabilities — positioning itself as the next-generation execution engine for Spark analytics.

References

• Apache Spark Documentation
• Gluten Project Repository
• Velox Documentation
• TPC-DS Benchmark
• “Velox: Meta's Unified Execution Engine” — VLDB 2022 Proceedings