PieRank: Embedded Large-Scale Sparse Matrix Processing
Michelle S. Zhou · 2023
I introduce PieRank, a library aimed at embedded large-scale sparse matrix processing. It enjoys significant advantages over previous state-of-the-art for handling big, sparse data sets in scalability, speed, and usability. My approach is also more general, allowing a wider variety of applications beyond graphs. Inspired by Raspberry Pi(e) for its low cost, I name the library PieRank with the goal of dramatically reducing the hardware requirements for cutting-edge sparse matrix processing. My experiments show PieRank outperforms GraphChi by a wide margin, making it possible for a single embedded device to meet or exceed the scalability and performance of server-class computers on big matrices such as AGATHA 2015, a deep-learning network with 5. 8B parameters as the largest instance in the popular SuiteSparse Matrix Collection. Available as an open-source project on GitHub under the permissive MIT license, PieRank compiles and runs on Linux, Mac OS, and Raspberry Pi OS.