Yusheng Zheng

AI Agent Infrastructure · Training & Inference Systems · Production at Scale · OSDI / ATC / ASPLOS

Yusheng Zheng云微

Ph.D. Candidate in Computer Science and Engineering at UC Santa Cruz

Building programmable and observable runtimes for AI infrastructure.

My research spans GPU runtime and policy mechanisms, distributed-training performance diagnosis, Linux extensibility, and system support for autonomous agents. I expect to graduate in 2027. I am currently a Student Researcher at ByteDance Seed, working on large-scale AI training infrastructure, observability, and systems performance. My ByteDance research periods are June–August 2026 and November 2026–present.

Expected Ph.D. 2027 • Santa Cruz · San Francisco Bay Area

OSDI ’25
First-author systems paper
10K+
OSS stars across original projects
350+
Research citations

Research

Selected Systems

Systems are the primary unit of my work: each combines a concrete mechanism, an implementation, and evaluation or downstream use.

SysOM-AI

research design & prototype

Continuous cross-layer diagnosis for production AI training, combining CPU stack profiling, GPU kernel tracing, and NCCL instrumentation with less than 0.4% overhead.

Validated by Alibaba collaborators across 80K+ GPUs; 94 confirmed production issues diagnosed.

bpftime

Creator / lead maintainer · OSDI 2025

A userspace eBPF runtime and verified application-extension framework combining verification, isolation, and dynamic instrumentation without application recompilation.

1.5K+ GitHub stars, active external users, and downstream integrations in the eBPF ecosystem.

gpu_ext

ASPLOS 2027

Safe and Programmable OS-Level GPU Resource Management with eBPF. A programmable policy runtime spanning the Linux GPU driver and device for scheduling, memory, and observability.

Up to 4.8× throughput improvement and 2× lower tail latency on evaluated AI workloads.

BPF-Ext

ATC 2026

Safely Extending the eBPF Compilation Pipeline with Native Operations. Extends compilation with native operations while retaining verification of the corresponding eBPF instruction sequences.

Up to 24% faster eBPF microbenchmarks and 12% faster production applications.

AgentSight

Creator / lead maintainer · agent runtime observability

System-level boundary tracing that correlates LLM and agent activity with process, file, network, and OS effects without requiring application SDK instrumentation.

Less than 3% average overhead; integrated into Alibaba Cloud Linux 4 Agentic Edition (ANOLISA).

Writing

Recent Notes

Research notes, systems experiments, and occasional writing outside computer systems.

Community

Presentations

Talk slides and posters on eBPF, AI agents, runtimes, and systems infrastructure.

OpenClaw

OpenClaw Open-Source AI Agents: Personal Assistants, Security, and Open Research Questions