University of Pennsylvania
Incoming Ph.D. student in Computer and Information Science.
Jiexiao Xu
I study computer systems. My current work asks how microservices can hold their end-to-end latency targets (SLOs) when a cluster is pushed past its capacity.
My research reaches across systems and graphics: I have built scalable, accurate crash-consistency testing for storage applications and optimized rendering pipelines for 3D Gaussian Splatting.
I have collaborated with Shihang Li, Simon Peter, Ratul Mahajan, Gilbert Bernstein, Yile Gu, and Baris Kasikci.
Education
University of Washington
Combined B.S./M.S. student in Computer Science.
News
- Starting a Ph.D. in Computer and Information Science at the University of Pennsylvania.
- Paper accepted to OOPSLA 2025 on application-level crash-consistency testing.
- Working on SLO-aware scheduling in microservices with Shihang Li, Simon Peter, and Ratul Mahajan.
- Exploring 3D Gaussian Splatting rendering optimizations with Gilbert Bernstein.
Research
SLO-aware microservice scheduling
When a cluster runs past capacity, microservice chains miss their end-to-end latency targets because each local scheduler is blind to downstream demand. I am building schedulers that propagate end-to-end deadlines and estimate downstream work, so the system sheds and prioritizes requests intelligently — raising the goodput an overloaded deployment can sustain within its SLO.
Scalable and accurate crash-consistency testing
Application-level crash-consistency testing blows up combinatorially: the number of crash states to check grows faster than any tool can keep up with. We cluster behaviorally similar crash states and test one representative per group, which makes testing tractable on real storage applications while still surfacing the consistency bugs an exhaustive search would find.
Paper ↗
doi:10.1145/3720431
Faster 3D Gaussian Splatting rendering
Real-time 3D Gaussian Splatting spends much of its per-frame budget sorting splats by depth. I explored rendering-pipeline optimizations that cut this sorting overhead, reducing runtime and memory use for real-time 3D reconstruction.