Broader Impacts of this Project

Large datasets in terabytes or petabytes are increasingly common, calling for new kinds of scalable machine learning approaches. While state-of-the-art techniques often use complex designs, specialized methods to store and work with large datasets, this project proposes a minimalist approach that forgoes such complexities, by leveraging the fundamental virtual memory capability found on all modern operating systems, to load into the virtual memory space the large datasets that are otherwise too large to fit in the computer's main memory. This main idea will allow developers to easily work with large datasets as if they were in-memory data, enabling them to create machine learning software that is significantly easier to develop and maintain, yet faster and more scalable. Developers will achieve higher work efficiency and make fewer programming errors; companies will reduce operating costs; and researchers will innovate methodology without getting bogged down by implementation details and scalability concerns. The proposed ideas could make a far-reaching impact on industry and academia, in science, education, and technology, as they face increasing challenges in applying machine learning on large datasets. The proposed ideas will also help train the next generation of scientists and engineers by allowing students to learn to work with large datasets in significantly simpler ways. As virtual memory is universally available on modern devices and operating systems, the proposed ideas will also work on mobile, low-power devices, enabling them to perform computation at unprecedented scales and speed.

Research Goals

This project investigates a fundamental, radical way to scale up machine learning algorithms based on virtual memory, one that may be easier to code and maintain, but currently under-utilized in by both single-machine and multi-machine distributed approaches. This research aims to develop deep understanding of this radical idea, its benefits and limitations, and to what extent these results apply in various settings, with respect to datasets, memory sizes, page sizes (e.g., from the default 4KB to the jumbo 2MB pages that enable terabyes of virtual memory space), and architectures (e.g., testing on distributed shared memory file systems like Lustre that support paging and virtual memory over large computer clusters). The researchers will build on their preliminary work on graph algorithms that already demonstrates significant speed-up over state-of-the-art approaches; they will extend their approach to a wide range of machine learning and data mining algorithms. They will also develop mathematical models and systematic approaches to profile and predict algorithm performance and energy usage based on extensive evaluation across platforms, datasets, and languages.