Jongseok Park

• Ph.D. Student at University of California, Berkeley

• Contact: js_park at berkeley dot edu 

• CV: Link (Last Updated Jan. 2025)

• Advisor: Prof. Ion Stoica and Prof. Alvin Cheung

Hello! I'm Jongseok Park, a Computer Science Ph.D. student in Electrical Engineering and Computer Sciences (EECS) at UC Berkeley, advised by Prof. Ion Stoica and Prof. Alvin Cheung. I'm part of the Sky Computing Lab. My interest lies in designing computation systems, scheduling algorithms, and acceleration hardware that enable the efficient use of new, high-cost applications, such as machine learning, by leveraging the characteristics of the target environment.

Check out the Gallery and Personal Projects for more of my past activities!

Overview

ASPEN is a novel parallel computation system for DNNs that allows fine-grained dynamic execution of DNNs. ASPEN removes synchronization barriers of tensor operators and expresses DNNs in dataflow graphs of fine-grained tiles, exposing novel computation opportunities across operators. ASPEN exploits these opportunities in runtime by dynamically locating and scheduling them in a distributed and asynchronous manner.

ASPEN enables opportunistic parallelism, a new class of parallelism for DNNs that is unavailable in operator-based approaches. ASPEN’s graph-wide parallel scheduling scope paired with dynamic, distributed runtime enables highly efficient resource utilization, scalability, and load balancing. Also, ASPEN achieves higher data reuse and reduced communications by letting each resource asynchronously traverse depthwise in the DNN graph.

Links & Materials

• [OpenReview]    [Paper]    [Github]    [Poster]    [Presentation Deck]    [Graphics (.ai file)]    [NeurIPS'23 Page (Requires Login)] 

Overview

mGEMM is a convolution computation algorithm that focuses on the low-latency requirements of mobile AI applications. mGEMM removes the memory duplication and reordering overhead (im2col) of existing matrix multiplication (GEMM) based convolution algorithms. mGEMM achieves this by adding filter spatial dimensions of convolutions in the assembly-level GEMM computation kernel, allowing the reuse of ARM NEON vector registers over GEMM iterations on different filter spatial dimensions.

mGEMM achieves improved latency, memory usage, and energy consumption in real-world mobile devices, and can be used as a drop-in replacement for existing convolution kernels. 

Links & Materials

• [MobiSys'22 Presentation (Youtube)]    [Paper]    [Github]    [Poster (GTS ver.)]    [Presentation Deck]    [Graphics (.pptx file)]