Intel® AVX-512 Accelerates PostgreSQL 18 CRC32C Checksums
Presented by:
Eshe Pickett
Eshé Pickett is a Cloud Software Development Engineer on Intel Corporation's Cloud Ecosystem Engineering team, where she specializes in performance analysis and optimizations leveraging Intel Architecture. Since joining Intel in 2005, Eshé has evolved from IT infrastructure specialist to system software architect, focusing on cloud-native technologies across enterprise environments. Eshé's team actively contributes to the PostgreSQL open-source community, with upstreamed patches in PostgreSQL 17 and 18, and pending submissions in the PostgreSQL 19 commitfest. Her recent work focuses on hardware-accelerated database optimizations, particularly leveraging Intel SIMD instructions. She holds a Master of Science in Computer Science from the University of Chicago and a Bachelor's degree from Northwestern University. Eshé's technical presentations have been featured at Intel's Design & Test Technology Conference and Software Professionals Conference. She was recognized in the Phoenix Business Journal's 40 Under 40 in 2017 and served on the Grace Hopper Conference leadership team from 2009-22.
No video of the event yet, sorry!
Abstract
PostgreSQL backup operations are critical for database reliability, but CRC32C checksum calculations create performance bottlenecks that scale with database size. This talk demonstrates how Intel AVX-512 vectorization in PostgreSQL 18's CRC32C implementation delivers measurable real-world performance improvements, reducing backup times by 11-14% with significant cost savings for enterprise deployments.
The Challenge: Traditional scalar CRC32C implementations process data sequentially, creating throughput limitations during backup operations. As database sizes grow, checksum calculation time becomes increasingly problematic, directly impacting backup windows and operational efficiency.
Our Solution: Our team contributed optimized CRC32C implementations to PostgreSQL 18, leveraging Intel AVX-512 SIMD instructions with 512-bit parallel processing capabilities. Building on the MIT-licensed fast-crc32c project, our implementation utilizes VPCLMULQDQ instructions on Intel® 4th Generation Xeon® Scalable processors to process multiple data streams simultaneously, improving throughput over previous Intel SSE4.2 implementations.
Test Configuration: We validated performance using dual Intel® Xeon® Platinum 8480+ processors (56 cores, 512GB DDR5-4800) running PostgreSQL 18RC1 on Ubuntu 22.04.5 with kernel 6.8.0. The optimization requires Intel SSE4.2 and PCLMUL support (Intel Core i7/i5 2nd gen+, Xeon 5500+) with optimal performance on Intel® 3rd Generation Xeon® Scalable processors or newer.
Results & Cost Impact: Testing across 15GB to 256GB databases showed consistent 11-14% performance improvements. A 256GB backup improved from 173.0 to 152.6 seconds (20.4 seconds saved). For enterprise scenarios with multiple daily backups, these savings compound significantly. Consider a self-managed PostgreSQL deployment on AWS EC2 (m7i.8xlarge at ~$1.73/hour): reducing backup windows by 20+ seconds per operation across multiple daily backups can save substantial compute costs, particularly for organizations with strict backup SLAs requiring dedicated backup windows.
Technical Implementation: The AVX-512 optimization processes 512 bits per instruction cycle compared to 128 bits in SSE4.2, with benefits scaling with dataset size. This architectural advantage becomes increasingly valuable for enterprise backup scenarios where absolute time savings matter most.
Session Agenda (50 minutes):
- Problem Context & Team Introduction (5m)
- Intel SIMD Architecture & AVX-512 Capabilities (10m)
- PostgreSQL CRC32C Implementation Deep-dive (15m)
- Benchmark Methodology & Configuration Details (10m)
- Real-world Results & Cost Analysis (7m)
- Community Impact & Future Directions (3m)
Speaker Bios: Eshe Pickett and Kelly McKeighan are Cloud Software Development Engineers on Intel Corporation's Cloud Ecosystem Engineering team, specializing in performance analysis and optimizations leveraging Intel Architecture for cloud workloads. Their team has actively contributed to the PostgreSQL community with upstreamed patches in PostgreSQL 17 and 18, with additional submissions pending in the PostgreSQL 19 commitfest.
Audience Value: Attendees will gain practical insights into hardware-accelerated database optimizations, understanding implementation details, configuration requirements, and quantifiable business impact for production PostgreSQL deployments.
- Date:
- Duration:
- 50 min
- Room:
- Conference:
- Postgres Conference: 2026
- Language:
- Track:
- Essentials
- Difficulty:
- Medium