Dr. Xipeng Shen

Modern machine learning programs are often written in Python, with the main computations specified through calls to some highly optimized libraries (e.g., TensorFlow, PyTorch). Existing performance optimizers focus on the static computation graphs specified by library APIs, but leave the influence from the hosting Python code largely unconsidered. This work proposes a new approach named HARP to address the problem. Through a set of novel techniques, HARP enables holistic analysis that spans across computation graphs and their hosting Python code. Refactoring based on HARP gives 1.3--3X and 2.07X average speedups on a set of TensorFlow and PyTorch programs. Read more...

This work proposes a new defensive technique for memory, especially useful for long-living objects on Non-Volatile Memory (NVM), or called Persistent Memory objects (PMOs). The method takes a distinctive perspective, trying to reduce memory exposure time by largely shortening the overhead in attaching and detaching PMOs into the memory space. It does it through a novel idea, embedding page table subtrees inside PMOs. The new technique reduces memory exposure time by 60% with a 5% time overhead (70% with 10.9% overhead). It allows much more frequent address randomizations (shortening the period from seconds to less than 41.4us), offering significant potential for enhancing memory security. Details to be posted in our ASPLOS'2020 paper.

Convolution Neural Networks (CNN) pruning is an important method to adapt a large CNN model attained on general datasets to a more specialized task or to fit a device with stricter space or power constraints. Finding the best pruned network is time-consuming. This work tackles the problem by creating a compiler-based framework named Wootz, which for the first time enables composability-based CNN pruning. Wootz shortens the state-of-art pruning process by up to 117.9X while producing significantly better pruning results. See our PLDI'2019 paper for more...

We propose the first known solution to enable direct document analytics on compressed data, which save 90.8% storage space and 77.5% memory usage, while halving the analytics time. It employs a hierarchical compression algorithm to convert analytics problems into graph traversal problems. The article presents a set of guidelines and assistant software modules for developers to effectively apply compression-based direct processing. See our VLDB'18, ICS'18 and ICDE'20 papers for more.

It is now an exciting time in computer systems research. New technologies such as machine learning and the Internet of Things (IoT) are rapidly enabling new capabilities that were only dreamed of a few years ago. At the same time, technology discontinuities such as the end of Moore’s Law and the inescapable Energy Wall combine with new challenges in security and privacy, and the rise of Artificial Intelligence (AI). Against this backdrop, an NSF-sponsored community visioning workshop convened about 150 researchers of multiple computer systems areas during ASPLOS'2018. The goal was to outline a few high-priority areas where inter-disciplinary research is likely to have a high payoff in the next 10 years. This report summarizes the workshop’s findings. (ACM DL link here.)