Tencent Cloud Database Kernel Research Accepted at SIGMOD and ICDE: Spitfire Buffer Manager and TLBtree Index

Today, two of Tencent Cloud's database kernel research projects were selected for inclusion in the prestigious SIGMOD and ICDE conferences, which are among the top three international database conferences.

The SIGMOD paper, titled "Spitfire: A Three‑Tier Buffer Manager for Volatile and Non‑Volatile Memory," introduces a multi‑level buffer manager that keeps hot pages in DRAM, warm pages in NVM, and uses a low‑cost hot‑data identification method to improve overall system performance and cost‑effectiveness.

Spitfire focuses on page migration strategies across three cache layers, proposing a probability‑based framework that determines where a page should reside—whether it should move from SSD to NVM or DRAM, and where evicted pages should be written. The authors also present a simulated‑annealing machine‑learning technique to automatically tune migration policies for arbitrary workloads and storage hierarchies.

By leveraging NVM's byte‑addressability and persistence, Spitfire can directly read/write NVM pages or dirty pages from DRAM without flushing them to SSD, reducing the cost of database recovery protocols, especially for write‑intensive workloads. This technology has been validated in TDSQL‑C (formerly CynosDB), Tencent's cloud‑native relational database compatible with MySQL and PostgreSQL.

The SIGMOD reviewers highlighted the importance of managing multi‑level buffers that include NVM and praised Spitfire as an innovative solution not previously explored in depth.

The ICDE paper, titled "TLBtree: A Read/Write‑Optimized Tree Index for Non‑Volatile Memory," presents a two‑layer persistent B+‑tree architecture. The upper layer is a read‑optimized, pointer‑free multi‑branch tree stored in a contiguous array, while the lower layer is a write‑optimized structure designed for NVM that ensures durability without a write‑ahead log.

TLBtree’s concurrent version employs lock‑free reads to minimize read‑write conflicts in multi‑process environments. Reviewers praised the novelty of combining multiple data structures to achieve read‑write balance on NVM and confirmed the experimental results.

These achievements continue Tencent Cloud's tradition of publishing high‑impact database research, building on prior collaborations such as the SIGMOD‑19 paper on AI‑driven automatic database tuning. The joint research center with Huazhong University of Science and Technology has produced over 20 papers and 14 patents, reinforcing Tencent's commitment to advancing foundational database and storage technologies.