Browsing by Author "Li, Yang"

Now showing 1 - 2 of 2
  • Thumbnail Image
    GPU-based Private Information Retrieval for On-Device Machine Learning Inference
    Lam, Maximilian; Johnson, Jeff; Xiong, Wenjie; Maeng, Kiwan; Gupta, Udit; Li, Yang; Lai, Liangzhen; Leontiadis, Ilias; Rhu, Minsoo; Lee, Hsien-Hsin S.; Reddi, Vijay Janapa; Wei, Gu-Yeon; Brooks, David; Suh, Edward (ACM, 2024-04-27)
    On-device machine learning (ML) inference can enable the use of private user data on user devices without revealing them to remote servers. However, a pure on-device solution to private ML inference is impractical for many applications that rely on embedding tables that are too large to be stored on-device. In particular, recommendation models typically use multiple embedding tables each on the order of 1-10 GBs of data, making them impractical to store on-device. To overcome this barrier, we propose the use of private information retrieval (PIR) to efficiently and privately retrieve embeddings from servers without sharing any private information. As off-the-shelf PIR algorithms are usually too computationally intensive to directly use for latency-sensitive inference tasks, we 1) propose novel GPU-based acceleration of PIR, and 2) co-design PIR with the downstream ML application to obtain further speedup. Our GPU acceleration strategy improves system throughput by more than 20× over an optimized CPU PIR implementation, and our PIR-ML co-design provides an over 5× additional throughput improvement at fixed model quality. Together, for various on-device ML applications such as recommendation and language modeling, our system on a single V100 GPU can serve up to 100, 000 queries per second—a > 100× throughput improvement over a CPU-based baseline—while maintaining model accuracy.
  • Thumbnail Image
    Modeling liquid droplet impact on a micropillar-arrayed viscoelastic surface via mechanically averaged responses
    Li, Yang; Cheng, Jiangtao (Taylor & Francis, 2023)
    Droplet impact on a substrate is an intriguing phenomenon that widely exists in our daily life and a broad range of industrial processes. However, droplet impact dynamics on soft textured surfaces are less explored and the underlying mechanisms remain elusive. Here, we report numerical simulation of droplet impact dynamics on a micropillar-arrayed soft surface using BASILISK, which involves a multiscale geometric domain containing the micropillars and droplet that are in the order of mu m and mm, respectively. As such, the volume of fluid (VOF) method is coupled with the finite volume method (FVM) to build the fluid fields and track their interface. From a conceptual point of view, the micropillared substrate is formed by imposing interstitial gaps into the otherwise intact soft material, whose viscoelastic properties can be quantified by gap density epsilon. Via a five-parameter generalized Maxwell model, the viscoelastic properties of the micropillared substrate can be approximated by its equivalent elastic response in the Laplace-Carson (LC) space, and the averaged bulk strain of the micropillared substrate in the real space is obtained by the inverse LC transform. Moreover, through parametric studies of splash extent, it turns out that for a specific epsilon, the splash is dramatically intensified with increasing impact velocity U-i. The splash also turns more violent with increasing ambient pressure P-a, which is evidenced by a larger splash angle of 114.44 degrees between the ejected sheet and the horizontal substrate at 5 atm. Conversely, the splash becomes more depressed with increasing surface tension sigma. Overall, the splash magnitudes of our simulations agree well with those predicted by the Kelvin-Helmholtz instability theory. By leveraging the LC transform in the fluid-viscoelastic solid interactions, our simulation methodology captures the main features of droplet impact dynamics on microstructured viscoelastic surfaces by means of the mechanically averaged responses while avoiding the predicament of domain scale inconsistency.