Canfor
50-Minute Talk
PostgreSQL was originally architected for an era of Uniform Memory Access (UMA) and modest CPU core counts. As the industry shifts toward high-density, multi-socket servers featuring 128+ cores, non-linear NUMA topologies, and emerging CXL memory tiers, the traditional "flat" shared memory model faces structural limitations. While continuous community optimizations have been highly effective, the overhead of maintaining global state is approaching a point of diminishing returns on modern hardware. This session analyzes the systemic bottlenecks that emerge at extreme scale—such as cacheline contention on global locks and the interconnect fabric strain during massive parallel scans—and argues that data locality must evolve from an optimization hint into a core architectural concern.
This community proposal explores optimizing PostgreSQL for modern hardware by partitioning global state structures to minimize synchronization overhead, investigating tiered buffer management for CXL-enabled memory hierarchies, and leveraging adaptive scheduling to better utilize heterogeneous CPU resources. By decoupling synchronization from global memory, the talk explores how the kernel can maintain its core identity while scaling to next-generation hardware.
Finally, we’ll look at a phased approach—starting with extension hooks and moving toward core locking refactors. The idea isn't to rewrite Postgres, but to find a practical way for the existing engine to handle the hardware we’re actually seeing in the data center today.
