Zejia Lin

Zejia Lin

Auto-generated profile, based on publicly available information

Zejia Lin

Sun Yat-Sen University, Guangzhou, China

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
22 March 2026
Verified
Conference Paper
Open access
ZLZejia Lin
Zejia Lin
HXHongxin Xu
Hongxin Xu
GCGuanyi Chen
Guanyi Chen
ZCZhiguang Chen
Zhiguang Chen
LYLu Yu-Tong
Lu Yu-Tong
XZXianwei Zhang
Xianwei Zhang
Modern large language model (LLM) serving systems confront inefficient GPU utilization due to the fundamental mismatch between compute-intensive prefill phase and memory-bound decode phase. While current practices attempt to address this by organizing these phases into hybrid batches, such solutions create an inefficient tradeoff that sacrifices either throughput or latency, leaving substantial GPU resources underutilized. For this, we identify two key root causes: 1) the prefill phase suffers from suboptimal compute utilization due to wave quantization and attention bottlenecks, and 2) hybrid batching disproportionately prioritizes latency over throughput, wasting both compute resources and memory bandwidth. To mitigate the issues, we present Bullet, a novel spatial-temporal orchestration system that eliminates these inefficiencies through fine-grained phase coordination. Bullet enables concurrent execution of prefill and decode requests, while dynamically provisioning GPU resources based on real-time performance modeling. By integrating SLO-aware scheduling and adaptive resource allocation, Bullet maximizes GPU utilization without compromising latency targets. Experimental evaluations on real-world workloads demonstrate that Bullet delivers 1.26× average throughput gains (up to 1.55×) over state-of-the-arts, while consistently meeting latency constraints.


20 January 2025
Verified
Conference Paper
Open access
MXMengyue Xi
Mengyue Xi
TGTianyu Guo
Tianyu Guo
XHXuanteng Huang
Xuanteng Huang
ZLZejia Lin
Zejia Lin
XZXianwei Zhang
Xianwei Zhang
Graphics Processing Units (GPUs) are essential for general-purpose applications and are commonly leveraging multi-level caches to alleviate memory access pressure. However, the default cache management may lose opportunities for optimal performance in different applications. Although existing cache bypassing techniques tend to address this challenge, these methods predominantly concentrate on single-level cache, thus restricting their potential for further enhancements. To mitigate this issue, we propose Mpache, a novel software-based mechanism designed to bypass multi-level caches based on the characterization of load instructions. Mpache constructs an interaction graph and analyzes the cooperation and contention among instructions. Then, the profiling data of bypassing effectiveness guides Mpache to select the appropriate cache levels to bypass for each instruction. Finally, the design is integrated into the compiler to enable automatic bypassing for existing workloads. Evaluations on off-the-shelf GPUs show that Mpache achieves an average 1.15× speedup over the default cache policy, and effectively outperforms prior arts.

15 December 2025 in ACM Transactions on Architecture and Code Optimization
Verified
Journal Article
Open access
WPWenxuan Pan
Wenxuan Pan
ZLZejia Lin
Zejia Lin
JDJiangsu Du
Jiangsu Du
XZXianwei Zhang
Xianwei Zhang
Nowadays, Graphics Processing Units (GPUs) dominate in a wide spectrum of computing realms with massive parallel processing capabilities. However, as resources are continuously integrated into GPUs, traditional serial execution often leads to underutilization. Prior studies have shown that allowing multiple kernels to run concurrently and share GPU resources can effectively improve both resource utilization and system throughput. Nonetheless, existing methods for solely automating concurrent kernel execution either schedule kernels within individual applications (i.e., kernel-level) or enable task concurrency across multiple applications (i.e., task-level), thus leaving substantial GPU capacity underexploited. Moreover, they inevitably introduce new programming frameworks, which incur cumbersome manual efforts and further impose substantial programming burdens on developers. To address these limitations, we propose HuntKTm , a hybrid scheduling and automatic management method that cooperates kernel-level and task-level concurrency to enhance system throughput with minimal code modification. Specifically, HuntKTm comprises a stream scheduler to assign kernels, a task scheduler to dispatch tasks onto GPUs, and a memory manager to reduce memory footprint. The stream scheduler applies a strategy to dispatch kernels to hardware queues and adopts a novel algorithm to remove redundant synchronizations in computational flow. Then, the task scheduler automatically issues tasks based on resource requirements and availability to support GPU sharing among uncooperative applications. Finally, the memory manager reduces memory footprint for tasks by limiting the lifetimes of memory objects, which thereby enables more tasks to execute simultaneously. Experimental results demonstrate that HuntKTm improves system throughput by 33.2% over the existing state-of-the-art CASE framework on a single machine equipped with four NVIDIA A100 GPUs and reaches 13.8% higher application-level performance over Taskflow, even with lessened programming efforts.