Amelie Chi Zhou
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Amelie Chi Zhou has a total of 69 co-authors. Here are the top co-authors by number of shared publications:
- Yusen Li
Nankai University
4 shared publications - Sitian Chen
Hong Kong Baptist University
3 shared publications - Xiaowen Chu
The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China
3 shared publications - Zhenheng Tang
Hong Kong University of Science and Technology
3 shared publications - Haobin Tan
College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, China
2 shared publications - Junlin Huang
Hong Kong University of Science and Technology (Guangzhou)
2 shared publications
Latest Publications
- Open Access
- 1 November 2026
- Published byĀ Institute of Electrical and Electronics Engineers (IEEE)Ā inĀ IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol.Ā 45 Ā (11) , 6290-6303
GPAQuant: Accelerating Personalized Recommendation With Group-Wise and PIM-Aware Quantization
Deep Learning Recommendation Models (DLRMs) have gained popularity in recommendation systems due to their effectiveness in handling large-scale recommendation tasks. However, the embedding layers of DLRMs have emerged as a significant performance bottleneck due to their high demands on memory capacity and bandwidth. Existing solutions, such as offloading to CPU memory, are hampered by communication overheads, while many proposed Processing-In-Memory (PIM) architectures are not yet commercially viable, limiting their practical impact. A promising approach, demonstrated by recent work like UpDLRM, is to accelerate these layers using commercial Processing-In-Memory (PIM) hardware. However, this approach does not address the challenge of ever-growing model sizes in industrial-scale systems. To tackle both memory bandwidth and capacity limitations, this paper proposes GPAQuant, a framework that co-designs PIM-based acceleration with model quantization. GPAQuant strategically stores quantized embedding tables (EMTs) in distributed DPU memory and offloads both the memory-intensive lookups and reductions to the DPUs. To address DPU-specific challenges, we propose four key innovations: (1) architecture-aware EMT partitioning that optimizes data placement to minimize inter-DPU communication, (2) frequency-aware workload balancing to evenly distribute access patterns across DPUs, (3) fine-grained EMT rearrangement that reduces quantization errors while preserving spatial locality, and (4) group-wise quantization to balance dequantization overhead and model accuracy. Extensive evaluations across six datasets and four DLRM architectures demonstrate GPAQuantās superiority: it achieves 4.6Ć speedup on embedding layers and 3.4Ć end-to-end inference acceleration compared to state-of-the-art CPU-GPU implementations and PIM baselines, with negligible accuracy loss.
- 1 November 2026
- Published byĀ Institute of Electrical and Electronics Engineers (IEEE)Ā inĀ IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol.Ā 45 Ā (11) , 6290-6303
Open AccessGPAQuant: Accelerating Personalized Recommendation With Group-Wise and PIM-Aware Quantization
Deep Learning Recommendation Models (DLRMs) have gained popularity in recommendation systems due to their effectiveness in handling large-scale recommendation tasks. However, the embedding layers of DLRMs have emerged as a significant performance bottleneck due to their high demands on memory capacity and bandwidth. Existing solutions, such as offloading to CPU memory, are hampered by communication overheads, while many proposed Processing-In-Memory (PIM) architectures are not yet commercially viable, limiting their practical impact. A promising approach, demonstrated by recent work like UpDLRM, is to accelerate these layers using commercial Processing-In-Memory (PIM) hardware. However, this approach does not address the challenge of ever-growing model sizes in industrial-scale systems. To tackle both memory bandwidth and capacity limitations, this paper proposes GPAQuant, a framework that co-designs PIM-based acceleration with model quantization. GPAQuant strategically stores quantized embedding tables (EMTs) in distributed DPU memory and offloads both the memory-intensive lookups and reductions to the DPUs. To address DPU-specific challenges, we propose four key innovations: (1) architecture-aware EMT partitioning that optimizes data placement to minimize inter-DPU communication, (2) frequency-aware workload balancing to evenly distribute access patterns across DPUs, (3) fine-grained EMT rearrangement that reduces quantization errors while preserving spatial locality, and (4) group-wise quantization to balance dequantization overhead and model accuracy. Extensive evaluations across six datasets and four DLRM architectures demonstrate GPAQuantās superiority: it achieves 4.6Ć speedup on embedding layers and 3.4Ć end-to-end inference acceleration compared to state-of-the-art CPU-GPU implementations and PIM baselines, with negligible accuracy loss.
- 1 October 2026
- Published byĀ Institute of Electrical and Electronics Engineers (IEEE)Ā inĀ IEEE Internet of Things Journal
- Vol.Ā 13 Ā (20) , 47358-47369
Stale Information Matters: Efficiently Tackling Data Heterogeneity in Asynchronous Federated Learning With Model Calibration
- 1 October 2026
- Published byĀ Institute of Electrical and Electronics Engineers (IEEE)Ā inĀ IEEE Internet of Things Journal
- Vol.Ā 13 Ā (20) , 47358-47369
Stale Information Matters: Efficiently Tackling Data Heterogeneity in Asynchronous Federated Learning With Model Calibration
- Open Access
- 4 September 2026
- Published byĀ Association for Computing Machinery (ACM)Ā inĀ ACM Transactions on Architecture and Code Optimization
- Vol.Ā 23 Ā (4) , 1-24
HARMONI: Heterogeneity-Aware I/O Scheduling for Mixed Workloads in SSD-Based HPC Systems
Modern HPC systems increasingly rely on tiered storage architectures with SSDs serving as a critical performance tier. However, the inherent asynchronous I/O characteristics of SSDs, including read/write bandwidth asymmetry and interference, pose significant challenges for traditional I/O schedulers. These challenges are exacerbated by the convergence of bursty HPC write workloads (e.g., checkpointing) and sustained AI read workloads (e.g., data streaming) on shared SSD infrastructure. Existing schedulers fail to adequately address these combined workloads, leading to suboptimal resource utilization. This article introduces HARMONI, a heterogeneity-aware reinforcement learning scheduler for mixed I/O in HPC storage systems. HARMONI leverages a graph neural network (GNN) to encode task-SSD dependencies and a hybrid interference predictor to adapt to hardware and I/O variations. Experimental results across diverse HPC and AI workloads demonstrate that HARMONI significantly reduces average makespan by up to 90% compared with state-of-the-art schedulers, effectively bridging the gap between evolving storage hardware and the dynamic I/O demands of modern HPC systems.
- 4 September 2026
- Published byĀ Association for Computing Machinery (ACM)Ā inĀ ACM Transactions on Architecture and Code Optimization
- Vol.Ā 23 Ā (4) , 1-24
Open AccessHARMONI: Heterogeneity-Aware I/O Scheduling for Mixed Workloads in SSD-Based HPC Systems
Modern HPC systems increasingly rely on tiered storage architectures with SSDs serving as a critical performance tier. However, the inherent asynchronous I/O characteristics of SSDs, including read/write bandwidth asymmetry and interference, pose significant challenges for traditional I/O schedulers. These challenges are exacerbated by the convergence of bursty HPC write workloads (e.g., checkpointing) and sustained AI read workloads (e.g., data streaming) on shared SSD infrastructure. Existing schedulers fail to adequately address these combined workloads, leading to suboptimal resource utilization. This article introduces HARMONI, a heterogeneity-aware reinforcement learning scheduler for mixed I/O in HPC storage systems. HARMONI leverages a graph neural network (GNN) to encode task-SSD dependencies and a hybrid interference predictor to adapt to hardware and I/O variations. Experimental results across diverse HPC and AI workloads demonstrate that HARMONI significantly reduces average makespan by up to 90% compared with state-of-the-art schedulers, effectively bridging the gap between evolving storage hardware and the dynamic I/O demands of modern HPC systems.
- Open Access
- 20 March 2026
- Published byĀ Association for Computing Machinery (ACM)Ā inĀ ACM Transactions on Architecture and Code Optimization
- Vol.Ā 23 Ā (1) , 1-25
PIMFuse: Kernel Fusion for Processing-in-Memory
Recent studies have demonstrated that processing-in-memory (PIM) can significantly accelerate memory-intensive applications across various domains. However, based on our experiments, we discover that PIM kernels in applications inherently have diverse resource requirements. To meet these resource requirements and fully utilize PIM resources, we first attempt to introduce the kernel fusion method to fuse PIM kernels with complementary resource requirements during the execution of multiple kernels. Based on our detailed study about PIM kernel fusion, we also reveal that both the kernel fusion combination and the thread allocation schemes during kernel fusion significantly impact overall performance. To determine the kernel fusion combination and thread allocation schemes that approach optimal performance, we propose PIMFuse, a framework leveraging kernel fusion to execute multiple PIM kernels. PIMFuse relies on accurate PIM kernel duration prediction models and fusion models to efficiently execute PIM kernels. Our experiments on real PIM devices show that PIMFuse reduces kernel execution time by up to 26.63% and on average by 15.51% compared to the baseline, and PIMFuse can be integrated with existing complex PIM applications to improve their performance. PIMFuse is publicly available at https://github.com/FanYang98/PIMFuse
- 20 March 2026
- Published byĀ Association for Computing Machinery (ACM)Ā inĀ ACM Transactions on Architecture and Code Optimization
- Vol.Ā 23 Ā (1) , 1-25
Open AccessPIMFuse: Kernel Fusion for Processing-in-Memory
Recent studies have demonstrated that processing-in-memory (PIM) can significantly accelerate memory-intensive applications across various domains. However, based on our experiments, we discover that PIM kernels in applications inherently have diverse resource requirements. To meet these resource requirements and fully utilize PIM resources, we first attempt to introduce the kernel fusion method to fuse PIM kernels with complementary resource requirements during the execution of multiple kernels. Based on our detailed study about PIM kernel fusion, we also reveal that both the kernel fusion combination and the thread allocation schemes during kernel fusion significantly impact overall performance. To determine the kernel fusion combination and thread allocation schemes that approach optimal performance, we propose PIMFuse, a framework leveraging kernel fusion to execute multiple PIM kernels. PIMFuse relies on accurate PIM kernel duration prediction models and fusion models to efficiently execute PIM kernels. Our experiments on real PIM devices show that PIMFuse reduces kernel execution time by up to 26.63% and on average by 15.51% compared to the baseline, and PIMFuse can be integrated with existing complex PIM applications to improve their performance. PIMFuse is publicly available at https://github.com/FanYang98/PIMFuse
- Open Access
- 15 November 2025
- Published byĀ Association for Computing Machinery (ACM)
- p. 789-804
UpANNS: Enhancing Billion-Scale ANNS Efficiency with Real-World PIM Architecture
- 15 November 2025
- Published byĀ Association for Computing Machinery (ACM)
- p. 789-804
Open AccessUpANNS: Enhancing Billion-Scale ANNS Efficiency with Real-World PIM Architecture
- Top Cited PapersOpen Access
- 3 August 2025
- Published byĀ Association for Computing Machinery (ACM)
- p. 5831-5841
BurstGPT: A Real-World Workload Dataset to Optimize LLM Serving Systems
- 3 August 2025
- Published byĀ Association for Computing Machinery (ACM)
- p. 5831-5841
Top Cited PapersOpen AccessBurstGPT: A Real-World Workload Dataset to Optimize LLM Serving Systems
- 3 April 2025
- Published byĀ Institute of Electrical and Electronics Engineers (IEEE)Ā inĀ IEEE Transactions on Parallel and Distributed Systems
- Vol.Ā 36 Ā (6) , 1161-1174
Distributed and Adaptive Partitioning for Large Graphs in Geo-Distributed Data Centers
Graph partitioning is of great importance to optimizing the performance and cost of geo-distributed graph analytics applications. However, it is non-trivial to obtain efficient and effective partitioning due to the challenges brought by the large graph scales, dynamic graph changes and the network heterogeneity in geo-distributed data centers (DCs). Existing studies usually adopt heuristic-based methods to achieve fast and balanced partitioning for large graphs, which are not powerful enough to address the complexity in our problem. Further, graph structures of many applications can change at various frequencies. Dynamic partitioning methods usually focus on achieving low latency to quickly adapt to changes, which unfortunately sacrifices partitioning effectiveness. Also, such methods are not aware of the dynamicity of graphs and can over sacrifice effectiveness for unnecessarily low latency. To address the limitations of existing studies, we propose DistRLCut, a novel graph partitioner which leverages Multi-Agent Reinforcement Learning (MARL) to solve the complexity of the partitioning problem. To achieve fast partitioning for large graphs, DistRLCut adapts MARL to a distributed implementation which significantly accelerates the learning process. Further, DistRLCut incorporates two techniques to trade-off between partitioning effectiveness and efficiency, including local training and agent sampling. By adaptively tuning the number of local training iterations and the agent sampling rate, DistRLCut is able to achieve good partitioning results within an overhead constraint required by graph dynamicity. Experiments using real cloud DCs and real-world graphs show that, compared to state-of-the-art static partitioning methods, DistRLCut improves the performance of geo-distributed graph analytics by 11%-95%. DistRLCut can partition over 28 million edges per second, showcasing its scalability for large graphs. With varying graph changing frequencies, DistRLCut can improve the performance by up to 71% compared to state-of-the-art dynamic partitioning.
- 3 April 2025
- Published byĀ Institute of Electrical and Electronics Engineers (IEEE)Ā inĀ IEEE Transactions on Parallel and Distributed Systems
- Vol.Ā 36 Ā (6) , 1161-1174
Distributed and Adaptive Partitioning for Large Graphs in Geo-Distributed Data Centers
Graph partitioning is of great importance to optimizing the performance and cost of geo-distributed graph analytics applications. However, it is non-trivial to obtain efficient and effective partitioning due to the challenges brought by the large graph scales, dynamic graph changes and the network heterogeneity in geo-distributed data centers (DCs). Existing studies usually adopt heuristic-based methods to achieve fast and balanced partitioning for large graphs, which are not powerful enough to address the complexity in our problem. Further, graph structures of many applications can change at various frequencies. Dynamic partitioning methods usually focus on achieving low latency to quickly adapt to changes, which unfortunately sacrifices partitioning effectiveness. Also, such methods are not aware of the dynamicity of graphs and can over sacrifice effectiveness for unnecessarily low latency. To address the limitations of existing studies, we propose DistRLCut, a novel graph partitioner which leverages Multi-Agent Reinforcement Learning (MARL) to solve the complexity of the partitioning problem. To achieve fast partitioning for large graphs, DistRLCut adapts MARL to a distributed implementation which significantly accelerates the learning process. Further, DistRLCut incorporates two techniques to trade-off between partitioning effectiveness and efficiency, including local training and agent sampling. By adaptively tuning the number of local training iterations and the agent sampling rate, DistRLCut is able to achieve good partitioning results within an overhead constraint required by graph dynamicity. Experiments using real cloud DCs and real-world graphs show that, compared to state-of-the-art static partitioning methods, DistRLCut improves the performance of geo-distributed graph analytics by 11%-95%. DistRLCut can partition over 28 million edges per second, showcasing its scalability for large graphs. With varying graph changing frequencies, DistRLCut can improve the performance by up to 71% compared to state-of-the-art dynamic partitioning.
- Open Access
- 1 March 2025
- Published byĀ Association for Computing Machinery (ACM)
- p. 1-7
ParaCoder: Parallel Code Generation with Large Language Model
- 1 March 2025
- Published byĀ Association for Computing Machinery (ACM)
- p. 1-7
Open AccessParaCoder: Parallel Code Generation with Large Language Model
- Top Cited PapersOpen Access
- 10 February 2025
- Published byĀ Association for Computing Machinery (ACM)Ā inĀ Proceedings of the ACM on Management of Data
- Vol.Ā 3 Ā (1) , 1-28
Tribase: A Vector Data Query Engine for Reliable and Lossless Pruning Compression using Triangle Inequalities
Approximate Nearest Neighbor Search (ANNS) is a critical problem in vector databases. Cluster-based index is utilized to narrow the search scope of ANNS, thereby accelerating the search process. Due to its scalability, it is widely employed in real-world vector search systems. However, existing cluster-based indexes often suffer from coarse granularity, requiring query vectors to compute distances with vectors of varying quality, thus increasing query complexity. Existing work aim to represent vectors with minimal cost, such as using product quantization (PQ) or linear transformations, to speed up ANNS. However, these approaches do not address the coarse granularity inherent in cluster-based index. In this paper, we present an efficient vector data query engine to enhance the granularity of cluster-based index by carefully subdividing clusters using diverse distance metrics. Building on this refined index, we introduce techniques that leverage triangle inequalities to develop highly optimized and distinct search strategies for clusters and vectors of varying qualities, thereby reducing the overhead of ANNS. Extensive experiments demonstrate that our method significantly outperforms existing in-memory cluster-based indexing algorithms, achieving up to an impressive 10Ć speedup and a pruning ratio exceeding 99.4%.
- 10 February 2025
- Published byĀ Association for Computing Machinery (ACM)Ā inĀ Proceedings of the ACM on Management of Data
- Vol.Ā 3 Ā (1) , 1-28
Top Cited PapersOpen AccessTribase: A Vector Data Query Engine for Reliable and Lossless Pruning Compression using Triangle Inequalities
Approximate Nearest Neighbor Search (ANNS) is a critical problem in vector databases. Cluster-based index is utilized to narrow the search scope of ANNS, thereby accelerating the search process. Due to its scalability, it is widely employed in real-world vector search systems. However, existing cluster-based indexes often suffer from coarse granularity, requiring query vectors to compute distances with vectors of varying quality, thus increasing query complexity. Existing work aim to represent vectors with minimal cost, such as using product quantization (PQ) or linear transformations, to speed up ANNS. However, these approaches do not address the coarse granularity inherent in cluster-based index. In this paper, we present an efficient vector data query engine to enhance the granularity of cluster-based index by carefully subdividing clusters using diverse distance metrics. Building on this refined index, we introduce techniques that leverage triangle inequalities to develop highly optimized and distinct search strategies for clusters and vectors of varying qualities, thereby reducing the overhead of ANNS. Extensive experiments demonstrate that our method significantly outperforms existing in-memory cluster-based indexing algorithms, achieving up to an impressive 10Ć speedup and a pruning ratio exceeding 99.4%.
- Open Access
- 12 August 2024
- Published byĀ Association for Computing Machinery (ACM)
- p. 866-875
Bandwidth-Aware and Overlap-Weighted Compression for Communication-Efficient Federated Learning
Current data compression methods, such as sparsification in Federated Averaging (FedAvg), effectively enhance the communication efficiency of Federated Learning (FL). However, these methods encounter challenges such as the straggler problem and diminished model performance due to heterogeneous bandwidth and non-IID (Independently and Identically Distributed) data. To address these issues, we introduce a bandwidth-aware compression framework for FL, aimed at improving communication efficiency while mitigating the problems associated with non-IID data. First, our strategy dynamically adjusts compression ratios according to bandwidth, enabling clients to upload their models at a close pace, thus exploiting the otherwise wasted time to transmit more data. Second, we identify the non-overlapped pattern of retained parameters after compression, which results in diminished client update signals due to uniformly averaged weights. Based on this finding, we propose a parameter mask to adjust the client-averaging coefficients at the parameter level, thereby more closely approximating the original updates, and improving the training convergence under heterogeneous environments. Our evaluations reveal that our method significantly boosts model accuracy, with a maximum improvement of 13% over the uncompressed FedAvg. Moreover, it achieves a 3.37 Ć speedup in reaching the target accuracy compared to FedAvg with a Top-K compressor, demonstrating its effectiveness in accelerating convergence with compression. The integration of common compression techniques into our framework further establishes its potential as a versatile foundation for future cross-device, communication-efficient FL research, addressing critical challenges in FL and advancing the field of distributed machine learning.
- 12 August 2024
- Published byĀ Association for Computing Machinery (ACM)
- p. 866-875
Open AccessBandwidth-Aware and Overlap-Weighted Compression for Communication-Efficient Federated Learning
Current data compression methods, such as sparsification in Federated Averaging (FedAvg), effectively enhance the communication efficiency of Federated Learning (FL). However, these methods encounter challenges such as the straggler problem and diminished model performance due to heterogeneous bandwidth and non-IID (Independently and Identically Distributed) data. To address these issues, we introduce a bandwidth-aware compression framework for FL, aimed at improving communication efficiency while mitigating the problems associated with non-IID data. First, our strategy dynamically adjusts compression ratios according to bandwidth, enabling clients to upload their models at a close pace, thus exploiting the otherwise wasted time to transmit more data. Second, we identify the non-overlapped pattern of retained parameters after compression, which results in diminished client update signals due to uniformly averaged weights. Based on this finding, we propose a parameter mask to adjust the client-averaging coefficients at the parameter level, thereby more closely approximating the original updates, and improving the training convergence under heterogeneous environments. Our evaluations reveal that our method significantly boosts model accuracy, with a maximum improvement of 13% over the uncompressed FedAvg. Moreover, it achieves a 3.37 Ć speedup in reaching the target accuracy compared to FedAvg with a Top-K compressor, demonstrating its effectiveness in accelerating convergence with compression. The integration of common compression techniques into our framework further establishes its potential as a versatile foundation for future cross-device, communication-efficient FL research, addressing critical challenges in FL and advancing the field of distributed machine learning.