TDB
Tensor-centric execution for SQL, subgraph, and graph analytics over GPU and heterogeneous runtimes.
TDB is a tensor-centric query execution direction for SQL, subgraph queries, and graph analytic workloads over GPU and heterogeneous accelerator runtimes.
Problem
Database workloads contain irregular operators — joins, aggregation, variable-length data processing, graph traversal, and subgraph matching — that do not map directly to the uniform tensor operators used by modern accelerator runtimes. This creates a gap between database execution and accelerator-native computation.
Core idea
TDB explores how to encode relational and graph workloads as tensor programs over runtimes such as PyTorch and TensorFlow, while preserving database-style execution semantics. The goal is to reuse optimized accelerator stacks without building a separate engine for each hardware backend.
My role
Research and system design across tensor-centric SQL and graph execution, including operator tensorization, graph topology encoding, compression, and out-of-XPU-memory execution.
Evidence
Impact
The TDB line shows that database and graph workloads can benefit from tensor runtimes and heterogeneous hardware. Reported results include 9.6× over TQP, 27.9× over HeavyDB, 12.2× over DuckDB on SQL workloads, and 50–100× GPU speedups for graph-query operators.