Data races are notorious bugs. They introduce non-determinism in programs behavior, complicate programs semantics, making it challenging to debug parallel programs. To make parallel programming easier, efficient data race detection has been a research topic in the last decades. However, existing data race detectors either sacrifice precision or incur high overhead, limiting their application to real-world applications and scenarios. This dissertation proposes approaches to improve the performance of dynamic data race detection without undermining precision, by identifying and removing metadata redundancy dynamically. This dissertation also explores ways to make it practical to detect data races dynamically for GPU programs, which has a disparate programming and execution model from CPU workloads. Further, this dissertation shows how the structured synchronization model in GPU programs can simplify the algorithm design of data race detection for GPU, and how the unique patterns in GPU workloads enable an efficient implementation of the algorithm, yielding a high-performance dynamic data race detector for GPU programs.