Junyi (Kyle) Shu

404 Westwood Plaza

Engineering VI, Room 496

Los Angeles, CA 90095

I am currently a postdoc researcher at UCLA working with Prof. Harry Xu. My mission is to build highly efficient, scalable infrastructure for next-generation AI. I have earned my Ph.D. from Peking University advised by Prof. Xin Jin and Prof. Xuanzhe Liu. Before that, I graduated from UC Berkeley with a Bachelor’s degree in Applied Math and Computer Science.

Beyond academic research, my career so far is defined by bridging deep technical execution with cross-functional leadership. Having engineered systems at AWS and collaborated closely with diverse cloud infrastructure teams at Alibaba Cloud, I deeply understand how to align complex technologies and coordinate with experts across multiple domains. Complementing this technical expertise, I have hands-on leadership experience directing engineering teams and managing various functions of early-stage startups.

news

Sep 23, 2026	Firework: Efficient and Transparent DSM with Hardware-Coherent CXL Shared Memory is accepted by ASPLOS ’27.
Sep 19, 2026	CacheSched: Efficient LLM Request Scheduling for Hierarchical Prefix Caching is accepted by ATC ’26.
Sep 16, 2026	COSMOS: Taming Heterogeneous Resource Demands in Serverless Functions with Complementary Swarming and Monotonic Scale-up is accepted by IEEE Transactions on Networking.

selected publications

* = co-first author, # = corresponding author

OSDI

Prism: Cost-Efficient Multi-LLM Serving via GPU Memory Ballooning

Shan Yu, Yifan Qiao, Mingyuan Ma, Yangmin Li, Shuo Yang, Xinyuan Tong, Yang Wang, Zhiqiang Xie, Yuwei An, Shiyi Cao, Ke Bao, Deepak Vij, Xiaoning Ding, Yichen Wang, Qingda Lu, Zhong Wang, Gao Gao, Harry Xu#, Junyi Shu#, Jiarong Xing#, and Ying Sheng#

In 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI 26) 2026

Deployed in production by Tencent, Alibaba Cloud, Red Hat, and others at a scale of 10,000+ GPUs.

Abs PDF Code

Artifacts Available

Inference providers must maintain availability for many LLMs, including low-volume but essential models, making resource efficiency increasingly important as token prices fall. Analysis of production traces reveals a dynamic bursty-group pattern in which sets of models become active together and shift over time; existing space- and time-sharing approaches lack principled mechanisms to adapt to this variability, forcing trade-offs between SLO adherence and efficiency. We observe that elastic memory allocation can unify spatial and temporal sharing. Based on this insight, we have developed Prism, a memory-centric LLM co-serving framework that applies memory ballooning to reclaim memory across models and support both forms of sharing under a single scheme. Prism’s balloon driver, referred to as kvcached, has been open-sourced at https://github.com/ovg-project/kvcached, and deployed in production environments across 10K+ GPUs.
OSDI

Burstable Cloud Block Storage with Data Processing Units

Junyi Shu, Kun Qian, Ennan Zhai, Xuanzhe Liu, and Xin Jin#

In 18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24) 2024

Deployed in production by Alibaba Cloud at a large scale.

Abs PDF Slides

Cloud block storage (CBS) is a key pillar of public clouds. Today’s CBS distinguishes itself from physical counterparts (e.g., SSDs) by offering unique burst capability as well as enhanced throughput, capacity, and availability. We conduct an initial characterization of our CBS product, a globally deployed cloud block storage service at public cloud provider Alibaba Cloud. A key observation is that the storage agent (SA) running on a data processing unit (DPU) which connects user VMs to the backend storage is the major source of performance fluctuation with burst capability provided. In this paper, we propose a hardware-software co-designed I/O scheduling system BurstCBS to address load imbalance and tenant interference at SA. BurstCBS exploits high-performance queue scaling to achieve near-perfect load balancing at line rate. To mitigate tenant interference, we design a novel burstable I/O scheduler that prioritizes resource allocation for base-level usage while supporting bursts. We employ a vectorized I/O cost estimator for comprehensive measurements of the consumed resources of different types of I/Os. Our evaluation shows that BurstCBS reduces average latency by up to 85% and provides up to 5× throughput for base-level tenants under congestion with minimal overhead. We verify the benefits brought by BurstCBS with a database service that internally relies on CBS, and show that up to 83% latency reduction is observed on customer workloads.
ASPLOS

Disaggregated RAID Storage in Modern Datacenters

Junyi Shu, Ruidong Zhu, Yun Ma, Gang Huang#, Hong Mei, Xuanzhe Liu, and Xin Jin#

In Proceedings of the 28th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 3 2023

Abs PDF Code Poster Slides

Artifacts Available Artifacts Functional Results Reproduced

RAID (Redundant Array of Independent Disks) has been widely adopted for decades, as it provides enhanced throughput and redundancy beyond what a single disk can offer. Today, enabled by fast datacenter networks, accessing remote block devices with acceptable overhead (i.e. disaggregated storage) becomes a reality (e.g., for serverless applications). Combining RAID with remote storage can provide the same benefits while creating better fault tolerance and flexibility than its monolithic counterparts. The key challenge of disaggregated RAID is to handle extra network traffic generated by RAID, which can consume a vast amount of NIC bandwidth. We present dRAID, a disaggregated RAID system that achieves near-optimal read and write throughput. dRAID exploits peer-to-peer disaggregated data access to reduce bandwidth consumption in both normal and degraded states. It employs non-blocking multi-stage writes to maximize inter-node parallelism, and applies pipelined I/O processing to maximize inter-device parallelism. We introduce bandwidth-aware reconstruction for better load balancing. We show that dRAID provides up to 3x bandwidth improvement. The results on a lightweight object store show that dRAID brings 1.5x-2.35x throughput improvement on various workloads.