Yanying Lin (林彦颖)

PhD Student of UCAS

Visiting at UPenn DSL

Email: yy.lin1@siat.ac.cn

Yanying Lin is a final-year PhD student in Computer Science at the University of Chinese Academy of Sciences (UCAS) with Prof. Kejiang Ye, currently visiting at the University of Pennsylvania DSL. His research focuses on Operating/Network Systems.

Research & Collaborations

My current research centers on serverless LLM systems and hardware acceleration, focusing on operating systems, and network performance to enhance large-scale model inference with efficient, reliable serving pipelines. Recently, I am particularly interested in disaggregated computing and communication architectures, dedicated to research on low-latency parallel communication and computing architectures.

I have had the privilege of collaborating with distinguished scholars in the system field: Prof. Rong Chen (SJTU), Prof. Xinyu Zhang (UCSD), Prof. Vincent Liu (UPenn).

At Penn, I focus on network/accelerator infrastructure for elastic inference. Previously at UCSD, I worked on FPGA-GPU acceleration for LLM systems. I also collaborate with Alibaba’s infrastructure and AIOS teams on industry-scale AI systems.

news

Oct 14, 2025	We released GenTD26, a production-level top-down dataset for serverless diffusion model serving! Learn more
Sep 10, 2025	Our paper “FlexPipe” has been accepted to EuroSys’26!

selected publications

EuroSys
FlexPipe: Adapting Dynamic LLM Serving Through Inflight Pipeline Refactoring in Fragmented Serverless Clusters

Yanying Lin, Shijie Peng, Chengzhi Lu, and 2 more authors

In Proceedings of the 21st European Conference on Computer Systems, 2026

Insight Bib HTML PDF

Insight: We find that for pipeline inference, the optimal online pipeline structure varies greatly depending on the incoming request distribution. Longer pipelines can absorb load spikes by dispersing demand, while shorter pipelines reduce communication overhead. Dynamically adjusting the pipeline structure according to the request distribution is an effective approach in serverless environments.
@inproceedings{lin2026flexpipe, title = {FlexPipe: Adapting Dynamic LLM Serving Through Inflight Pipeline Refactoring in Fragmented Serverless Clusters}, author = {Lin, Yanying and Peng, Shijie and Lu, Chengzhi and Xu, Chengzhong and Ye, Kejiang}, booktitle = {Proceedings of the 21st European Conference on Computer Systems}, year = {2026}, }
SoCC
Understanding Diffusion Model Serving in Production: A Top-Down Analysis of Workload, Scheduling, and Resource Efficiency

Yanying Lin, Shuaipeng Wu, Shutian Luo, and 8 more authors

In Proceedings of the 2025 ACM Symposium on Cloud Computing, 2025

Insight Bib PDF Slides

Insight: The main insights of this work are obtained from production systems. It focuses on the performance of online inference, middleware, and hardware layers, as well as the relationships between these three components. We find that skewed request distributions force systems to stress network and cache layers, resulting in poor resource locality and reduced system performance.
@inproceedings{lin2025understanding, title = {Understanding Diffusion Model Serving in Production: A Top-Down Analysis of Workload, Scheduling, and Resource Efficiency}, author = {Lin, Yanying and Wu, Shuaipeng and Luo, Shutian and Xu, Hong and Shen, Haiying and Ma, Chong and Shen, Min and Chen, Le and Xu, Chengzhong and Qu, Lin and others}, booktitle = {Proceedings of the 2025 ACM Symposium on Cloud Computing}, pages = {1--15}, year = {2025}, }
ICDCS
Quart: Latency-Aware FaaS System for Pipelining Large Model Inference

Yanying Lin, Yanbo Li, Shijie Peng, and 5 more authors

In Proceedings of the 44th IEEE International Conference on Distributed Computing Systems, 2024

Insight Bib PDF

Insight: We find that when using pipeline inference for LLMs, each stage exhibits significant load differences due to varying resource allocation and throughput across different workloads. For example, under stable or bursty traffic patterns, throughput disparities between upstream and downstream stages can create performance-critical bottlenecks in the pipeline.
@inproceedings{lin2024quart, title = {Quart: Latency-Aware FaaS System for Pipelining Large Model Inference}, author = {Lin, Yanying and Li, Yanbo and Peng, Shijie and Tang, Yingfei and Luo, Shutian and Shen, Haiying and Xu, Chengzhong and Ye, Kejiang}, booktitle = {Proceedings of the 44th IEEE International Conference on Distributed Computing Systems}, year = {2024}, }
IEEE TSC
Serving LLM in Distributed GPU Cluster with Fine-Grain Pipeline Constraints

Yanying Lin, Shijie Peng, Shuaipeng Wu, and 4 more authors

IEEE Transactions on Services Computing, 2025

Insight Bib

Insight: LLMs are increasingly deployed in distributed GPU clusters using pipeline parallelism for efficient inference. However, existing approaches treat the entire pipeline as a monolithic unit, resulting in coarse-grained resource management and suboptimal SLO compliance. Each pipeline stage has distinct workload characteristics, but current systems cannot leverage this for better performance control. We propose fine-grained pipeline constraints that manage each stage independently, enabling precise SLO allocation at the stage level.
@article{lin2025serving, title = {Serving LLM in Distributed GPU Cluster with Fine-Grain Pipeline Constraints}, author = {Lin, Yanying and Peng, Shijie and Wu, Shuaipeng and Li, Yanbo and Lu, Chengzhi and Ye, Kejiang and Xu, Chengzhong}, journal = {IEEE Transactions on Services Computing}, year = {2025}, }