Researchers tested the impact of Huge Pages (HP) on PostgreSQL 17 performance using a synthetic workload. The test environment included an Intel Xeon E5-2697 v2 processor and a virtual machine running Debian 13 with 8 cores and 100 GB of RAM. The database was configured with 40 GB of shared_buffers.
The study compared standard 4 KB memory pages against 2 MB Huge Pages. Using 2 MB pages reduced page faults from 1,861,651 to 1,044 over a 25-second test period. This configuration allowed the system to reduce kernel-space CPU cycles by 10.4% and increase user-space processing time by 5.6%.
In simple SELECT queries, Huge Pages provided an 8% increase in transactions per second (TPS). For complex queries involving JOINs, sorting, and aggregation, the performance gain dropped to 4.5%. The researchers noted that while absolute efficiency remains consistent, the relative gain diminishes as computational complexity increases.
Tests with 1 GB pages showed lower efficiency than 2 MB pages on the Ivy Bridge architecture due to hardware limitations in the L1 dTLB cache. The authors conclude that Huge Pages are most effective for systems with at least 50 GB of RAM, though they emphasize that real-world 1C application performance gains will likely be modest.
