Open Source Summit Europe 2025

Linux’s current writeback infrastructure, while robust, was designed before large folios, CXL-tiered memory, and AI workloads demanding low-latency, high-throughput I/O. Today, workloads like RAG pipelines using vector databases with buffered I/O, and memory tiering on CXL, are exposing scalability limits in how the kernel handles writeback.

This talk presents a forward-looking view on evolving Linux’s writeback model. We’ll explore how the single-threaded design stalls page migration and reduces memory compaction effectiveness—affecting hugepage allocations and folio movement across memory tiers, contributing to fragmentation. On the storage side, parallelizing writeback improves throughput and responsiveness under dirty-page pressure, especially for sustained-write workloads with large memory footprints on High capacity SSDs.

We’ll also touch on early experiments within the kernel community, including efforts to make writeback more filesystem-geometry aware and parallelize it based on overwrites/new allocations.

This session invites open source community to reimagine writeback as a scalable, performance-critical component in Linux.