Biography
Chang Guo
I am a PhD Candidate in Computer Science at Arizona State University, advised by Prof. Zhichao Cao. Before joining ASU, I earned a B.E. in Automation and a B.B.A. in Business Administration from Tsinghua University.
My research focuses on indexing, caching, and storage systems for LLM inference and scientific computing. LegoIndex received the ACM HPDC 2025 Best Student Paper Award.
I am currently a Software Engineering Intern on the NVIDIA Dynamo team. I work on tiered KV cache management across GPU, CPU, and SSD.
News
- Started as a Software Engineering Intern on the NVIDIA Dynamo team.
- O3-LSM appeared at ACM SIGMOD 2026.[paper]
- LegoIndex received the ACM HPDC 2025 Best Student Paper Award.[paper]
- Mosaic-Cache appeared at ACM HotStorage 2025.[paper]
- Started a research internship at ByteDance.
- Our work on ZNS SSDs for persistent caching appeared at ACM HotStorage 2024.[paper]
- CaaS-LSM appeared at ACM SIGMOD 2024.[paper]
- SAS-Cache and Prophet appeared at IEEE MSST 2024.[SAS-Cache][Prophet]
- Returned to Oak Ridge National Laboratory for a second research internship.
- Started a research internship with the ADIOS2 team at Oak Ridge National Laboratory.
- Started my PhD in Computer Science at Arizona State University.
Selected Research
Experience
Software Engineering Intern
Santa Clara, CA- Defined backend configuration and publisher-relative locality contracts for vLLM native KV offloading. Separated storage type from deployment topology while preserving backward compatibility.
- Built self-describing KV events across vLLM, Dynamo, and llm-d. Routed CPU and local storage placements into Dynamo's lower-tier KV index and made unsupported media fail closed.
- Deduplicated replicated MLA KV across tensor-parallel ranks in shared CPU memory. At TP=N, this provides about N× the effective offload capacity and reduces D2H store traffic by N×.
Selected PRs:[vLLM #48906][vLLM #48281][vLLM #48150][vLLM #43468][Dynamo #10368][Dynamo #11571][llm-d #680]
Research Intern
San Jose, CA- Co-developed GPU-centric storage for LLM checkpointing and KV cache offloading, integrated with LMCache and vLLM.
Research Intern
Oak Ridge, TN- Built client-side caching for remote scientific data, including partial-overlap reuse and metadata indexing.
- Developed Mosaic-Cache during the internship; the work appeared at ACM HotStorage '25 and was upstreamed through three merged ADIOS2 pull requests.
Paper:[HotStorage '25]Selected PR:[ADIOS2 #4210]
Algorithm Engineer
Beijing, China- Built NLP pipelines for information extraction, topic modeling, and sentiment analysis over financial documents and business news.
Education
PhD Candidate, Computer Science
Intelligent Data Infrastructure Lab; advisor: Prof. Zhichao Cao.
B.E. in Automation
Dual degree: B.B.A. in Business Administration.
Exchange Student
Selected Open-Source Contributions
vLLM
Architecture: Defined a publisher-relative locality contract for filesystem and object-store tiers. Established backend configuration boundaries and tier ownership for native KV offloading.
Placement events: Added self-describing events across CPU, filesystem, and object-store tiers so external routers can track offloaded KV blocks.
Shared CPU offloading & MLA KV deduplication: Moved native CPU offloading onto a shared mmap, eliminating allocator over-allocation. Enabled single-copy MLA KV across the default CPU and tiering backends; at TP=N, this provides about N× effective CPU capacity and reduces D2H store traffic by N×. Made secondary entries reusable across TP sizes.
[shared CPU backend][default CPU dedup][tiering MLA dedup][cross-TP secondary reuse]
NVIDIA Dynamo
Native offloading integration: Routed vLLM CPU and local storage KV events to Dynamo's lower-tier index. Rejected remote or unsupported events before they could pollute GPU routing and namespace state.
Routing controls & adoption: Exposed host and disk cache-hit weights and documented the supported vLLM native-offloading path.
llm-d
Improved KV-event compatibility and cache-index correctness for vLLM offloading.
ADIOS2
Built client-side caching for remote scientific-data reads, with partial-overlap reuse, dataset-aware keys, and query metadata.
Teaching
CSE 511: Data Processing at Scale
Teaching Assistant · Graduate-level
CSE 330: Operating Systems
Teaching Assistant · Undergraduate-level
Awards
- ACM HPDC 2025 Best Student Paper Award for LegoIndex.
- Student Travel Grants: USENIX FAST 2025, ACM HPDC 2025, and ACM HotStorage 2025.
Academic Service
- Shadow Program Committee: VLDB 2027, EuroSys 2026, and ACM SIGMETRICS 2026.
- Artifact Evaluation Committee: MLSys 2026.