Speeding Up MPC Query Execution through Efficient and Flexible Intermediate Result Size Trimming

Abstract:

There is a growing interest in Secure Collaborative Analytics using Secure Multi-Party Computation (MPC) in enterprises. Fully oblivious query execution, however, can be prohibitively expensive, in part due to the lack of data reduction between SQL operators. Recent works demonstrated that relaxing the protection of intermediate result sizes in an operator tree reduces overall computation and has a large performance benefit -- yet they offer no flexibility in navigating the performance/privacy trade-off. In our ongoing work, we propose a general and flexible method for accelerating Secure Analytics by reducing intermediate result sizes in a controlled manner. This is achieved through a lightweight operator inserted after any oblivious SQL operator that decreases the intermediate result size from fully oblivious toward the true size, based on configurable strategies. Our work enables navigating the performance/privacy trade-off on a per-operator and per-query basis – it also lays down the foundation for a future MPC query planner that can reason about performance and privacy targets when composing physical query plans.