Ming-Chang Yang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Ming-Chang Yang has a total of 131 co-authors. Here are the top co-authors by number of shared publications:
- Yuan-Hao Chang
National Taiwan University
14 shared publications - Chun-Feng Wu
Department of Computer Science, National Yang Ming Chiao Tung University, Hsinchu, Taiwan, Taiwan
11 shared publications - Yingjia Wang
The Chinese University of Hong Kong, Hong Kong, Hong Kong
7 shared publications - Shuo-Han Chen
National Yang Ming Chiao Tung University, Hsinchu, Taiwan
6 shared publications - Tsun-Yu Yang
The Chinese University of Hong Kong
6 shared publications - James Cheng
The Chinese University of Hong Kong, Hong Kong, China
5 shared publications
Latest Publications
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 7566-7576
RecCompl: Efficient Model Compilation for Industrial Scale Recommendation Models with PyTorch 2
Deep Learning Recommendation Models (DLRMs) play a key role to power real-world recommendation and ranking, yet their growing complexity has made production deployment increasingly challenging. While PyTorch 2 (PT2) offers promising performance and productivity gains through automated model compilation, its initial release lacked critical features needed for DLRM adoption. In this work, we present RecCompl, a comprehensive model compilation system that enables efficient model compilation of industrial scale DLRMs with PT2. RecCompl addresses key compatibility issues by extending operator coverage, minimizing graph breaks, avoiding unnecessary recompilation, and generalizing graph transformation. Besides, to meet the high requirement on model exploration, we introduce a configuration-based user interface that decouples compilation settings from model code, allowing fine-grained control without intrusive changes. Despite these improvements, efficiency gaps remain in achieving a production-ready compilation system. To close them, we introduce systematic designs and engineering optimizations that enhance compilation time, memory management, and online deployment reliability. RecCompl is widely adopted, and delivers up to 60% higher training throughput while achieving substantially lower compilation latency, consistent performance across varying memory budgets, and stable online deployment.
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 7566-7576
Open AccessRecCompl: Efficient Model Compilation for Industrial Scale Recommendation Models with PyTorch 2
Deep Learning Recommendation Models (DLRMs) play a key role to power real-world recommendation and ranking, yet their growing complexity has made production deployment increasingly challenging. While PyTorch 2 (PT2) offers promising performance and productivity gains through automated model compilation, its initial release lacked critical features needed for DLRM adoption. In this work, we present RecCompl, a comprehensive model compilation system that enables efficient model compilation of industrial scale DLRMs with PT2. RecCompl addresses key compatibility issues by extending operator coverage, minimizing graph breaks, avoiding unnecessary recompilation, and generalizing graph transformation. Besides, to meet the high requirement on model exploration, we introduce a configuration-based user interface that decouples compilation settings from model code, allowing fine-grained control without intrusive changes. Despite these improvements, efficiency gaps remain in achieving a production-ready compilation system. To close them, we introduce systematic designs and engineering optimizations that enhance compilation time, memory management, and online deployment reliability. RecCompl is widely adopted, and delivers up to 60% higher training throughput while achieving substantially lower compilation latency, consistent performance across varying memory budgets, and stable online deployment.
- Open Access
- 14 July 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Storage
Boosting ZNS SSD Write Performance with Zora
Zoned Namespace (ZNS) SSD presents a new class of storage devices that offers low cost, high and consistent performance, and software definability. Nevertheless, ZNS demands strict sequential writing within a zone, which severely limits the write performance of small-sized requests. Unfortunately, existing solutions are all sub-optimal due to unsatisfactory write performance, the need for invasive host software modifications, and/or compromised read performance. In this paper, we propose Zora , a host-SSD collaborative framework, to fully address the write performance limitation for ZNS SSDs. Zora includes two core components: a zone area abstraction tailored for generic writes and a request admission control module located in the Linux block layer. They collaborate to boost ZNS SSD write performance to near-optimal (i.e., up to 94.1% in raw SSD access, up to 96.2% in filesystem workloads, and 24.2% to 48.5% in applications such as Varmail, OLTP, and SQLite), while benefiting upper-level applications and filesystems with full transparency.
- 14 July 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Storage
Open AccessBoosting ZNS SSD Write Performance with Zora
Zoned Namespace (ZNS) SSD presents a new class of storage devices that offers low cost, high and consistent performance, and software definability. Nevertheless, ZNS demands strict sequential writing within a zone, which severely limits the write performance of small-sized requests. Unfortunately, existing solutions are all sub-optimal due to unsatisfactory write performance, the need for invasive host software modifications, and/or compromised read performance. In this paper, we propose Zora , a host-SSD collaborative framework, to fully address the write performance limitation for ZNS SSDs. Zora includes two core components: a zone area abstraction tailored for generic writes and a request admission control module located in the Linux block layer. They collaborate to boost ZNS SSD write performance to near-optimal (i.e., up to 94.1% in raw SSD access, up to 96.2% in filesystem workloads, and 24.2% to 48.5% in applications such as Varmail, OLTP, and SQLite), while benefiting upper-level applications and filesystems with full transparency.
- Open Access
- 8 May 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Storage
- Vol. 22 (4) , 1-30
Enabling Cost-effective and Efficient All-Distance Sketching in Out-of-core Approximate Graph System
The All-Distances Sketch (ADS) is a powerful and theoretically-sound sketching scheme that captures neighborhood information in graphs for approximate processing. It enables high-accuracy estimation of many useful applications with a guarantee of accuracy and can significantly accelerate the execution times by orders of magnitude. However, ADS requires a substantial amount of space that is multiple times larger than the graph data. More seriously, existing studies mainly focus on managing ADSs in memory, posing an increasing challenge for users who aim to leverage ADS for large-scale graph processing, particularly in light of the exponential growth of real-world graphs nowadays. To this end, this paper introduces Oasis, an O ut-of-core A pproximate graph SYS tem that brings the ADS technique into practical use by leveraging storage effectively. Specifically, Oasis offers a holistic framework that facilitates both ADS construction and estimation. For ADS construction, it allows users to adjust the memory usage based on the machine’s available memory and enable an efficient construction process. For ADS estimation, Oasis provides a user-friendly interface to easily execute the estimators while mitigating the impact of slow storage I/O. Evaluation results show that Oasis provides a practical graph processing solution with exceptional execution time and low memory usage, at the cost of a slight decrease in accuracy.
- 8 May 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Storage
- Vol. 22 (4) , 1-30
Open AccessEnabling Cost-effective and Efficient All-Distance Sketching in Out-of-core Approximate Graph System
The All-Distances Sketch (ADS) is a powerful and theoretically-sound sketching scheme that captures neighborhood information in graphs for approximate processing. It enables high-accuracy estimation of many useful applications with a guarantee of accuracy and can significantly accelerate the execution times by orders of magnitude. However, ADS requires a substantial amount of space that is multiple times larger than the graph data. More seriously, existing studies mainly focus on managing ADSs in memory, posing an increasing challenge for users who aim to leverage ADS for large-scale graph processing, particularly in light of the exponential growth of real-world graphs nowadays. To this end, this paper introduces Oasis, an O ut-of-core A pproximate graph SYS tem that brings the ADS technique into practical use by leveraging storage effectively. Specifically, Oasis offers a holistic framework that facilitates both ADS construction and estimation. For ADS construction, it allows users to adjust the memory usage based on the machine’s available memory and enable an efficient construction process. For ADS estimation, Oasis provides a user-friendly interface to easily execute the estimators while mitigating the impact of slow storage I/O. Evaluation results show that Oasis provides a practical graph processing solution with exceptional execution time and low memory usage, at the cost of a slight decrease in accuracy.
- Open Access
- 20 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 2066-2075
Hitcher: Efficient GPU-based Vector Search via Cluster-Centric Kernel and Hitch-Ride Ordering
Similarity-based vector search, which retrieves the most similar vectors to a given query vector from a large vector dataset, underlies many applications such as search, recommendation, and Large Language Models (LLMs). Some systems run vector search on GPUs to enjoy GPU's high parallelism, but we observe that they are limited in query throughput and latency. In particular, their query-centric GPU kernel conducts computation independently for each query, failing to reuse data loaded to the GPU shared memory across queries and leading to a low GPU compute utilization. While their batch-based task reordering rearranges computation for queries in a batch to reduce CPU-GPU data transfer, but latency is prolonged since each query needs to wait for its slowest task. To tackle these problems, we propose Hitcher. Specifically, to reuse data across queries and improve GPU utilization, Hitcher implements a cluster-centric GPU kernel to batch computation on the same data for multiple queries. To reduce query latency, Hitcher adopts the hitch-ride ordering, which preserves the arrival order for query processing while batching computation across queries to improve efficiency. Hitcher can also offload computation tasks to the CPU to reduce CPU-GPU data transfer and utilize multiple GPUs. Experimental results show that Hitcher achieves up to 22× lower P99 query latency and 9× higher query throughput when compared with the state-of-the-art GPU-based vector query processing systems.
- 20 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 2066-2075
Open AccessHitcher: Efficient GPU-based Vector Search via Cluster-Centric Kernel and Hitch-Ride Ordering
Similarity-based vector search, which retrieves the most similar vectors to a given query vector from a large vector dataset, underlies many applications such as search, recommendation, and Large Language Models (LLMs). Some systems run vector search on GPUs to enjoy GPU's high parallelism, but we observe that they are limited in query throughput and latency. In particular, their query-centric GPU kernel conducts computation independently for each query, failing to reuse data loaded to the GPU shared memory across queries and leading to a low GPU compute utilization. While their batch-based task reordering rearranges computation for queries in a batch to reduce CPU-GPU data transfer, but latency is prolonged since each query needs to wait for its slowest task. To tackle these problems, we propose Hitcher. Specifically, to reuse data across queries and improve GPU utilization, Hitcher implements a cluster-centric GPU kernel to batch computation on the same data for multiple queries. To reduce query latency, Hitcher adopts the hitch-ride ordering, which preserves the arrival order for query processing while batching computation across queries to improve efficiency. Hitcher can also offload computation tasks to the CPU to reduce CPU-GPU data transfer and utilize multiple GPUs. Experimental results show that Hitcher achieves up to 22× lower P99 query latency and 9× higher query throughput when compared with the state-of-the-art GPU-based vector query processing systems.
- Open Access
- 2 April 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (1) , 1-25
Making LSM-Tree-based Key-Value Store Practical and Efficient for Multi-Tenant Serverless Cloud Databases
Cloud databases revolutionize data processing and storage by providing on-demand and scalable services housed on the cloud infrastructure. Multi-tenancy and serverless are two key tenets transforming the cloud database architecture, offering significant cost savings and simplified user management. However, we found that cloud databases using LSM-tree-based key-value stores as the storage engine face a crucial conundrum when adopting this promising multi-tenant serverless architecture. Specifically, LSM-tree-based key-value store encounters a critical dilemma between maintaining performance service-level agreements (SLAs) for tenants and over-subscribing storage bandwidth for high cost-efficiency. In this paper, we present FlexEngine, a novel LSM-tree-based key-value store, which for the first time enables the practical and efficient adoption of LSM tree in multi-tenant serverless cloud databases. FlexEngine introduces a series of designs to navigate the above dilemma, including a two-level (i.e., partition- and node-level) I/O admission control framework and a two-stage compaction deferral mechanism. We implement FlexEngine on a commercially-deployed RocksDB and perform comprehensive experiments on both production traces and micro-level workloads. The experimental results demonstrate that FlexEngine can significantly improve the capability to over-subscribe storage bandwidth, which leads to high cost-efficiency, while still promising consistent performance SLAs for users.
- 2 April 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (1) , 1-25
Open AccessMaking LSM-Tree-based Key-Value Store Practical and Efficient for Multi-Tenant Serverless Cloud Databases
Cloud databases revolutionize data processing and storage by providing on-demand and scalable services housed on the cloud infrastructure. Multi-tenancy and serverless are two key tenets transforming the cloud database architecture, offering significant cost savings and simplified user management. However, we found that cloud databases using LSM-tree-based key-value stores as the storage engine face a crucial conundrum when adopting this promising multi-tenant serverless architecture. Specifically, LSM-tree-based key-value store encounters a critical dilemma between maintaining performance service-level agreements (SLAs) for tenants and over-subscribing storage bandwidth for high cost-efficiency. In this paper, we present FlexEngine, a novel LSM-tree-based key-value store, which for the first time enables the practical and efficient adoption of LSM tree in multi-tenant serverless cloud databases. FlexEngine introduces a series of designs to navigate the above dilemma, including a two-level (i.e., partition- and node-level) I/O admission control framework and a two-stage compaction deferral mechanism. We implement FlexEngine on a commercially-deployed RocksDB and perform comprehensive experiments on both production traces and micro-level workloads. The experimental results demonstrate that FlexEngine can significantly improve the capability to over-subscribe storage bandwidth, which leads to high cost-efficiency, while still promising consistent performance SLAs for users.
- Open Access
- 20 March 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (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 by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (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
- 16 December 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 22 (4) , 1-25
Pac-PIM: A Parallel Communication Framework for Commodity Processing-in-memory Systems
Processing-in-memory (PIM) is a promising paradigm to effectively alleviate the bottleneck of memory access for memory-intensive applications. UPMEM PIM-enabled DIMMs, the first commercially available PIM platform integrating processing elements (PEs) in DIMMs, exemplifies this trend. However, the communication overhead between the host and PIM-enabled DIMMs remains significant. Although previous studies have attempted to reduce this overhead between the host and PIM-enabled DIMMs, our detailed investigation reveals that, from the software perspective, the native communication framework fails to fully leverage its parallelism for communication. We propose to combine rank-level parallelism and process-level parallelism in the layered fashion to take advantage of these and avoid their respective drawbacks. And we revamp the native communication framework by integrating this combination to introduce Pac-PIM ( Pa rallel C ommunication Framework for PIM ). Our evaluation, conducted using real UPMEM PIM-enabled DIMMs and representative PIM-enabled applications, shows that Pac-PIM significantly improves communication performance in all scenarios, with an average speedup of 1.53× compared to the native communication framework with excellent scalability. In addition, Pac-PIM can further improve communication performance along with other optimizations.
- 16 December 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 22 (4) , 1-25
Open AccessPac-PIM: A Parallel Communication Framework for Commodity Processing-in-memory Systems
Processing-in-memory (PIM) is a promising paradigm to effectively alleviate the bottleneck of memory access for memory-intensive applications. UPMEM PIM-enabled DIMMs, the first commercially available PIM platform integrating processing elements (PEs) in DIMMs, exemplifies this trend. However, the communication overhead between the host and PIM-enabled DIMMs remains significant. Although previous studies have attempted to reduce this overhead between the host and PIM-enabled DIMMs, our detailed investigation reveals that, from the software perspective, the native communication framework fails to fully leverage its parallelism for communication. We propose to combine rank-level parallelism and process-level parallelism in the layered fashion to take advantage of these and avoid their respective drawbacks. And we revamp the native communication framework by integrating this combination to introduce Pac-PIM ( Pa rallel C ommunication Framework for PIM ). Our evaluation, conducted using real UPMEM PIM-enabled DIMMs and representative PIM-enabled applications, shows that Pac-PIM significantly improves communication performance in all scenarios, with an average speedup of 1.53× compared to the native communication framework with excellent scalability. In addition, Pac-PIM can further improve communication performance along with other optimizations.
- Open Access
- 12 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 169-185
Demeter: A Scalable and Elastic Tiered Memory Solution for Virtualized Cloud via Guest Delegation
- 12 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 169-185
Open AccessDemeter: A Scalable and Elastic Tiered Memory Solution for Virtualized Cloud via Guest Delegation
- Open Access
- 28 September 2025
- Published by Association for Computing Machinery (ACM)
- p. 1-10
Special Sessions - Design and Optimization for AI/ML Acceleration on Resource-constrained and Edge Systems
- 28 September 2025
- Published by Association for Computing Machinery (ACM)
- p. 1-10
Open AccessSpecial Sessions - Design and Optimization for AI/ML Acceleration on Resource-constrained and Edge Systems
- Open Access
- 26 September 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Embedded Computing Systems
- Vol. 24 (5s) , 1-25
Unlocking the Full Potential of Dual-Interface SSDs: A Comprehensive Hardware and Software Perspective
The legacy block interface for I/O benefits from data locality but faces challenges with I/O amplification due to the frequent small read-write operations common in most applications. Dual-Interface SSDs, which integrate block-interface Flash memory with byte-addressable memory, create opportunities for application redesign by reducing unnecessary read-write amplification on storage devices. However, current Dual-Interface SSD hardware remains limited in terms of size and functionality. Existing software designs for Dual-Interface SSDs often use the byte-addressable space as sequential logs with basic batch reclamation. While this space allows random access, and batch reclamation introduces significant tail latency caused by excessive read and write-back operations. To fully exploit the potential of Dual-Interface SSDs, we have developed a prototype on a hardware-software configurable platform. This publicly accessible Dual-Interface SSD offers realistic and optimized performance, overcoming the size and functionality limitations of previous designs. In addition, we demonstrate the ability of Dual-Interface SSDs to reduce write amplification in a traditional Copy-on-Write B-Tree data store. By employing two key techniques—Tree Pointer Relocation, which decouples indirection from the tree structure, and Tree Node Accommodation, which enables small-sized key-value pair updates to be processed directly in the byte-addressable space—we significantly improve the efficiency of storage operations. Our evaluation reveals that these techniques, when applied to Dual-Interface SSDs, achieve performance gains of 30.5% and 52.5% compared to a CoW B-Tree operating on traditional block-interface SSDs.
- 26 September 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Embedded Computing Systems
- Vol. 24 (5s) , 1-25
Open AccessUnlocking the Full Potential of Dual-Interface SSDs: A Comprehensive Hardware and Software Perspective
The legacy block interface for I/O benefits from data locality but faces challenges with I/O amplification due to the frequent small read-write operations common in most applications. Dual-Interface SSDs, which integrate block-interface Flash memory with byte-addressable memory, create opportunities for application redesign by reducing unnecessary read-write amplification on storage devices. However, current Dual-Interface SSD hardware remains limited in terms of size and functionality. Existing software designs for Dual-Interface SSDs often use the byte-addressable space as sequential logs with basic batch reclamation. While this space allows random access, and batch reclamation introduces significant tail latency caused by excessive read and write-back operations. To fully exploit the potential of Dual-Interface SSDs, we have developed a prototype on a hardware-software configurable platform. This publicly accessible Dual-Interface SSD offers realistic and optimized performance, overcoming the size and functionality limitations of previous designs. In addition, we demonstrate the ability of Dual-Interface SSDs to reduce write amplification in a traditional Copy-on-Write B-Tree data store. By employing two key techniques—Tree Pointer Relocation, which decouples indirection from the tree structure, and Tree Node Accommodation, which enables small-sized key-value pair updates to be processed directly in the byte-addressable space—we significantly improve the efficiency of storage operations. Our evaluation reveals that these techniques, when applied to Dual-Interface SSDs, achieve performance gains of 30.5% and 52.5% compared to a CoW B-Tree operating on traditional block-interface SSDs.