Human-Centered Performance Analysis

Execution visualization highlights time spent in GPU activity and waits. A memory-request, data-arrival, and task-continuation chain supports causal analysis and human-AI reasoning. A researcher compares designs and decides what to change, guiding the next analysis.

Research questions

How can a visualization connect execution details to the cause of a performance bottleneck?
Which views help architects move from an aggregate statistic to the events behind it?
How can an AI assistant help with performance analysis while keeping its conclusions verifiable?
How do experts use visual tools to compare and revise architecture designs?

Publications

2027

ASPLOS 2027

TritonParse: Multi-IR Provenance and Reproducible Debugging for Triton Kernel Compilation

Ying Li (co-first author), Yueming Hao (co-first author), David Berard, Xu Zhao, Barney Huang, Wenyuan Chi, Agron Tsai, Prithvi Patel, Srivatsan Ramesh, Manman Ren, Hongtao Yu, James Wu, Dan Zimmerman, Nicholas J Riasanovsky, Adam Mainz, Bert Maher, Peiying Hua, Daohang Shi, Taylor Robie, Yuanwei Fang, Shuofei Zhu, Shangdi Yu, Nikhil Patel, Bin Bao, Yang Chen, Jan Szczepaniec, Haishan Zhu, Patrick Li, Shane Nay, Adnan Aziz, Bill Yoshimi, Alexey Loginov, Yifan Sun

(Accepted; to appear.)