Master’s Thesis

Master’s Thesis

Discovery and design of new cuDF operators to support efficient Velox execution on GPUs

Ref. 2026_008

Velox is a high-performance, open-source execution engine for large-scale data processing workloads, including SQL query execution. It provides a modular operator-based execution model in which relational query plans are expressed as compositions of operators implementing primitive operations. Recently, GPU-accelerated data processing frameworks such as NVIDIA cuDF have demonstrated significant performance advantages; however, there remain gaps between the abstractions exposed by cuDF and the operator requirements of modern SQL engines.

The objective of this project is to systematically discover, design, and implement novel cuDF operators that better match the computational patterns arising in Velox query execution. The student will use workload-driven analysis and automated operator discovery tools such as OpenEvolve (or related techniques) to study SQL query plans and execution traces, identify recurring computation motifs, and propose new GPU-native operators. A central theme of the project is the exploration of operator abstraction levels, investigating questions such as: What constitutes a meaningful operator to be executed using one or more GPUs? How coarse or fine-grained should these operators be? How does operator granularity affect performance, generality, and maintainability?

The newly developed cuDF operators will be integrated into Velox through GPU-backed operator implementations, accompanied by comprehensive unit and integration testing. Their effectiveness will be evaluated using the industry-standard TPC-H benchmark.

The position is expected to take place at IBM Research in Zurich for the duration of one semester. The earliest start date is August 2026. Note that this position will be unpaid.

Requirements

• Strong C++ and Python proficiency;
• Experience with benchmarking and performance profiling;
• Basic knowledge of databases and SQL;
• Independent learning and working abilities.

Preferred experience/skills

• Experience with SQL engines, data processing systems, or GPU programming (CUDA, cuDF, RAPIDS, etc.);
• Familiarity with query execution, relational algebra, columnar data layouts, and memory management;
• Familiarity with query optimization, compiler techniques, or IR-based systems;
• Prior research experience or publications in systems, databases, or high-performance computing.

Expected project steps

• Study the Velox architecture and its operator execution model.
• Review cuDF primitives and their current usage patterns in SQL execution.
• Analyze SQL workloads and execution traces (TPC-H) to identify inefficiencies and missing abstractions, e.g., fallbacks to CPU execution, common sequences of operations.
• Discover recurring computation patterns and candidate operators.
• Systematically explore operator abstraction levels, defining and evaluating alternative operator boundaries.
• Design and implement new cuDF operators and integrate them into Velox.
• Develop unit and integration tests to ensure correctness and robustness.
• Evaluate performance using the TPC-H benchmark, analyzing throughput, kernel efficiency, and memory behavior.
• (stretch goal) Derive general operator design principles and prepare results for publication.

We actively encourage and support students to publish at top-tier conferences, e.g., NeurIPS, VLDB. We expect the output of this project to also result in an open-source contribution.

Diversity & Work Environment

IBM is committed to fostering diversity and inclusion in the workplace. You will join an open, multicultural research environment that values different perspectives and supports flexible working arrangements. Our goal is to help all genders and backgrounds thrive professionally while maintaining a healthy work-life balance.

How to Apply

If you are interested in this position, please submit your application through the button below.