BIFOLD Talk: Portable Query Compilation in SYCLDB and Autonomous Accelerator Specialization

Portable Query Compilation in SYCLDB and Autonomous Accelerator Specialization

Abstract:

Modern analytical and data-intensive engines need portable query execution frameworks that target heterogeneous hardware while maintaining high compute and memory efficiency.

We introduce SYCLDB, a portable database engine that uses Apache Calcite for query parsing and optimization, and translates the resulting query plans into hardware-accelerated operator pipelines using SYCL/AdaptiveCpp across CPUs and GPUs.

In this talk, we focus on our evaluation of standard analytical workloads, including SSB and TPC-H, demonstrating the benefits of SYCLDB's portable compilation model on GPUs. We further extend its execution model with JIT operator fusion and register-resident intermediate representations, reducing intermediate data movement and kernel launch overhead. However, achieving near-roofline performance requires additional hardware-specific optimizations. We address this through SHADB, an autonomous framework using profile-guided optimization (PGO) loop and low-level GPU performance counters to discover efficient physical execution strategies. Then, we explore which optimizations identified by SHADB, such as collision-free direct-mapped lookups and asymmetric thread coarsening, can be lifted into SYCLDB's runtime.

Finally, I will discuss how SYCLDB's compilation model and fused operator pipelines can be extended to stream processing across heterogeneous edge-to-cloud environments.

Bio:
Ivan Kabadzhov is a Ph.D. candidate in the Data Systems group at EURECOM and Sorbonne University, advised by Prof. Raja Appuswamy. His research is in the intersection of heterogeneous database systems, portable GPU query engines, and AI-driven kernel synthesis. During his doctoral studies, he also worked as a Software Engineering intern on C++ infra at Google/Chromium. Previously, he worked on profiling and scaling distributed dataframes in the ROOT team at CERN.