DataKernelBench：LLM 能否在 GPU 上优化数据库查询？

View PDF HTML (experimental)

Abstract:GPUs increasingly accelerate database systems, but query-specific peak performance still often relies on hand-written kernels. Existing LLM kernel benchmarks focus on machine learning operators, leaving irregular, heterogeneous, data-movement-heavy database-style operators untested. We introduce DataKernelBench, which translates SQL into validated PyTorch TorchPlan programs and evaluates LLMs that optimize either the core tensor-bounded snippet or the full query in CUDA or Triton through execution-guided repair. Across ten proprietary and open-weight models on TPC-H SF10 with an H100 GPU, the strongest full-query CUDA configuration achieves $2.11\times$ speedup over the TorchPlan baseline at full pass rate. We find that higher-performing implementations commonly use kernel fusion and execution-strategy changes, stronger models benefit most from full-query specialization, and workload context matters more than hardware context. To handle data larger than GPU memory, we extend TorchPlan with Dask-cuDF for on-demand partition loading on TPC-H SF100 with four H100 GPUs, achieving $2.54\times$ speedup.
Project page: this https URL

Comments:	Accepted at EMNLP 2026
Subjects:	Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Databases (cs.DB); Machine Learning (cs.LG); Programming Languages (cs.PL)
Cite as:	arXiv:2608.25061 [cs.CL]
	(or arXiv:2608.25061v3 [cs.CL] for this version)
	https://doi.org/10.48550/arXiv.2608.25061 arXiv-issued DOI via DataCite

Submission history

From: Gokul Karthik Kumar [view email]
[v1] Tue, 25 Aug 2026 18:57:39 UTC (2,340 KB)
[v2] Thu, 27 Aug 2026 17:07:30 UTC (2,340 KB)
[v3] Wed, 14 Oct 2026 21:33:09 UTC (2,340 KB)