Low Overhead Scalable Sorting for DNA Analysis using Apache Arrow Flight

Variant calling is a classic example of DNA data analysis, in which variants from DNA sequence data are identified. In recent years, a number of frameworks have been developed for implementing the variant calling pipeline. As a result, deeper scientific insights have been gained into genomics. From the computational point of view, however, these frameworks suffer from either excessive serialization and deserialization or lack of scalability. This work proposes using Apache Arrow as the unified in-memory data format and employing its associated communication protocol Flight to enable horizontal scalability for the sorting stage of a variant calling pipeline. Compared to other protocols, Arrow Flight features high throughput, low overhead, and compatibility with the Arrow format. For example, 43\% of the data shuffling time in a standalone distributed sorting application can be saved by replacing the original protocol with Arrow Flight. In addition, several techniques, including parallelization, hashing the keys for sorting to integers, and leveraging zero-copy conversion are proven to be useful for performance optimization under certain conditions. They together can reduce the execution time of the distributed sorting application from 71 seconds to only 12 seconds, achieving a 5.9x speedup. Linear scalability is also observed when the input data consists of integers keys. After integrating this sorting application into the variant calling pipeline, the performance of sorting in the pipeline outperforms the counterpart step in the state-of-the-art SparkGA2. Concretely, a speedup of 21x is achieved in the case of using 4 nodes and 10x is achieved in the case of using 12 nodes