Xinghang Yu · AI Infra & HPC

01 — About

About me

I'm a Computer Science student at Tianjin University, and I mostly work on infrastructure — the storage and systems that sit underneath machine learning.

Right now I'm at ByteDance, working on the Parameter Server — the high-throughput, low-latency store behind the AML machine-learning platform. Before that, competitive programming (I'm an ICPC silver medalist) gave me most of my data-structure and performance habits.

I like working across the whole stack, from low-level code up to the models on top. One project on astronomical light-curve classification grew into a first-author paper accepted by ApJS. Outside of systems, I'm into astronomy.

At a glance

FocusML Infra · Storage

StackC++ · CUDA · Distributed Storage

GitHubbestdo77

EducationCS · Tianjin University

LocationTianjin, China

InterestsSystems · Astronomy

02 — Experience & Education

My journey

2026.06 — Present

Infra Intern

ByteDance · Data Platform, AML-Engine

On the core infra team behind Parameter Server & KV-cache — the unified, stateful, high-availability parameter storage that powers AML, ByteDance's ML platform for recommendation, ads and search.
Core-path optimization: working on the Parameter Server hot path — the high-throughput, low-latency store for large-scale sparse features and large models across training and online serving.
Eviction & operability: building and tuning eviction strategy and day-to-day operability for a system that has to stay stable under real production load.

2026.03 — 2026.05

Infra Intern

National Supercomputing Center in Tianjin

Worked on LLM inference infrastructure — multi-LoRA batch scheduling on heterogeneous accelerators, enabling concurrent adapters in offline serving.
Tuned the vLLM serving path to overlap compute and transfer, cutting adapter-switch latency and GPU memory fragmentation.

2023.09 — Present

B.Eng. in Computer Science

Tianjin University (985/211)

Foundations in systems, parallel & heterogeneous computing, and algorithms — the base for the performance work I do today.

03 — Projects

Selected projects

Parallel Model Eviction for a Large-scale Parameter Server

ByteDance · Infra

2026 · C++ · Parameter Server · AOF + SSD tiered storage

Ordered concurrency: introduced per-region coordinators that keep the append-only log and the SSD write path strictly ordered while eviction runs across many threads.
~35% faster: parallelized the eviction hot path (single-thread falls back to the original code path as a safety net) — end-to-end eviction on a large production model sped up roughly 35% with near-flat memory and virtually no extra disk I/O.
Correctness first: verified by bit-exact, per-FID comparison of dump files across single-threaded, multi-threaded and concurrent-model runs — zero mismatches against the original path.

~35%faster eviction

1%extra memory

~0%extra disk I/O

0FID mismatches

Incremental Management & Real-time Classification of Large-scale Astronomical Light Curves

Lead · Accepted by ApJS

2025.03 — 2026.06 · TDengine · LightGBM · ONNX · C++ · Python

High-performance storage: designed a one-table-per-source model with HEALPix spatial indexing on TDengine — 48M rows ingested in 33s, millisecond cone searches (full scan 15s → 50ms).
Lightweight inference: replaced a 21.8GB random forest with LightGBM exported to ONNX — 87× smaller (252MB), memory from 22GB to <10MB, 93.6% accuracy with no GPU.
Incremental triggering: confidence filtering re-classifies sources as new data arrives; high-confidence results emit at 50% curve completeness to trigger follow-up.

48M/33singest rate

87×model shrink

93.6%accuracy

04 — Honors