Mingxing Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Mingxing Zhang has a total of 79 co-authors. Here are the top co-authors by number of shared publications:
- Yongwei Wu
Tsinghua University, Beijing, China
12 shared publications - Kang Chen
Tsinghua University, Beijing, China
7 shared publications - Jinlei Jiang
Tsinghua University, Beijing, China
6 shared publications - Yingdi Shan
Tsinghua University, Beijing, China
5 shared publications - Teng Ma
Alibaba Group, Hangzhou, Zhejiang, China
4 shared publications - Shaonan Ma
Qiyuan Lab, Beijing, China
3 shared publications
Latest Publications
- Open Access
- 11 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 16-29
DualPath: Accelerating Agentic LLM Inference by Harvesting Disaggregated KV-Cache Storage I/O
- 11 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 16-29
Open AccessDualPath: Accelerating Agentic LLM Inference by Harvesting Disaggregated KV-Cache Storage I/O
- Open Access
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
SSD Reforge: Dynamic Parity Management to Extend SSD Lifetime under Aging
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessSSD Reforge: Dynamic Parity Management to Extend SSD Lifetime under Aging
- Open Access
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
GWrite: Offloading Replication in Disaggregated Memory Systems
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessGWrite: Offloading Replication in Disaggregated Memory Systems
- Open Access
- 13 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 249-264
BCCE: Block-Centric GPU Co-Design for Real-Time Range-Top-K Query at Scale
- 13 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 249-264
Open AccessBCCE: Block-Centric GPU Co-Design for Real-Time Range-Top-K Query at Scale
- Open Access
- 18 May 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (3) , 1-27
CMANNS: GPU-Accelerated Graph Index Construction for ANNS via Compute–Memory Disaggregation
Graph-based approximate nearest neighbor search (ANNS) delivers state-of-the-art accuracy latency tradeoffs, yet index construction remains the bottleneck: fusing dense distance evaluation with irregular traversal or pruning collapses GPU throughput, and limited device memory forces costly data movement at scale. To address these problems, in this paper, we present CMANNS, a GPU-accelerated graph index construction framework that preserves the algorithmic rules of target graph (e.g., NSG and HNSW) and its query procedure. The core idea is compute–memory (CM) disaggregation : distance evaluation is reformulated as high–arithmetic-intensity GEMMs on Tensor Core accelerators with fused epilogues, while memory-intensive phases employ hot-set–aware on-chip locality (e.g., shared-memory staging, warp-cooperative gathers and scatters) to maximize effective bandwidth. To scale beyond the HBM capacity, we stream device-sized shards through a double-buffered pipeline and write back only compact adjacency. Data transfers and kernel execution overlap, so each shard completes in roughly the time of the slower step, keeping the GPU highly utilized even with irregular access. Across seven benchmarks, CMANNS reduces end-to-end index build time by up to 13.05x (vs. FAISS) and 2.20x (vs. FLASH), increases the cache hit rate by up to 58.7% , and preserves vector query latency and recall.
- 18 May 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (3) , 1-27
Open AccessCMANNS: GPU-Accelerated Graph Index Construction for ANNS via Compute–Memory Disaggregation
Graph-based approximate nearest neighbor search (ANNS) delivers state-of-the-art accuracy latency tradeoffs, yet index construction remains the bottleneck: fusing dense distance evaluation with irregular traversal or pruning collapses GPU throughput, and limited device memory forces costly data movement at scale. To address these problems, in this paper, we present CMANNS, a GPU-accelerated graph index construction framework that preserves the algorithmic rules of target graph (e.g., NSG and HNSW) and its query procedure. The core idea is compute–memory (CM) disaggregation : distance evaluation is reformulated as high–arithmetic-intensity GEMMs on Tensor Core accelerators with fused epilogues, while memory-intensive phases employ hot-set–aware on-chip locality (e.g., shared-memory staging, warp-cooperative gathers and scatters) to maximize effective bandwidth. To scale beyond the HBM capacity, we stream device-sized shards through a double-buffered pipeline and write back only compact adjacency. Data transfers and kernel execution overlap, so each shard completes in roughly the time of the slower step, keeping the GPU highly utilized even with irregular access. Across seven benchmarks, CMANNS reduces end-to-end index build time by up to 13.05x (vs. FAISS) and 2.20x (vs. FLASH), increases the cache hit rate by up to 58.7% , and preserves vector query latency and recall.
- Open Access
- 20 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 140-151
Faico: Faithful and Complete Knowledge Graph Augmented Reasoning
- 20 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 140-151
Open AccessFaico: Faithful and Complete Knowledge Graph Augmented Reasoning
- Open Access
- 2 April 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (1) , 1-28
From Prefix Cache to Fusion RAG Cache: Accelerating LLM Inference in Retrieval-Augmented Generation
Retrieval-Augmented Generation enhances Large Language Models by integrating external knowledge, which reduces hallucinations but increases prompt length. This increase leads to higher computational costs and longer Time to First Token. To mitigate this issue, existing solutions aim to reuse the preprocessed KVCache of each retrieved chunk to accelerate RAG. However, the lack of cross-chunk contextual information leads to a significant drop in generation quality, leaving the potential benefits of KVCache reuse largely unfulfilled. The challenge lies in how to reuse the precomputed KVCache chunk while preserving generation quality. We propose FusionRAG, a novel inference framework that optimizes both the preprocessing and reprocessing stages of RAG. In the offline preprocessing stage, we embed information from other related text chunks into each chunk, while in the online reprocessing stage, we recompute the KVCache for tokens that the model focuses on. As a result, we achieve a better trade-off between generation quality and efficiency. According to our experiments, FusionRAG significantly improves generation quality at the same recomputation ratio compared to previous state-of-the-art solutions. By recomputing fewer than 15% of the tokens, FusionRAG achieves up to 70% higher normalized-F1 scores than baselines and reduces TTFT by 2.66-9.39× compared to Full Attention.
- 2 April 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (1) , 1-28
Open AccessFrom Prefix Cache to Fusion RAG Cache: Accelerating LLM Inference in Retrieval-Augmented Generation
Retrieval-Augmented Generation enhances Large Language Models by integrating external knowledge, which reduces hallucinations but increases prompt length. This increase leads to higher computational costs and longer Time to First Token. To mitigate this issue, existing solutions aim to reuse the preprocessed KVCache of each retrieved chunk to accelerate RAG. However, the lack of cross-chunk contextual information leads to a significant drop in generation quality, leaving the potential benefits of KVCache reuse largely unfulfilled. The challenge lies in how to reuse the precomputed KVCache chunk while preserving generation quality. We propose FusionRAG, a novel inference framework that optimizes both the preprocessing and reprocessing stages of RAG. In the offline preprocessing stage, we embed information from other related text chunks into each chunk, while in the online reprocessing stage, we recompute the KVCache for tokens that the model focuses on. As a result, we achieve a better trade-off between generation quality and efficiency. According to our experiments, FusionRAG significantly improves generation quality at the same recomputation ratio compared to previous state-of-the-art solutions. By recomputing fewer than 15% of the tokens, FusionRAG achieves up to 70% higher normalized-F1 scores than baselines and reduces TTFT by 2.66-9.39× compared to Full Attention.
- Open Access
- 28 March 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Computer Systems
- Vol. 44 (3) , 1-39
TrEnv-X: Transparently Share Serverless Execution Environments Across Different Functions and Nodes
Serverless computing is renowned for its computation elasticity, yet its full potential is often constrained by the requirement for functions to operate within local and dedicated background environments, resulting in limited memory elasticity. To address this limitation, this paper introduces TrEnv-X , a co-designed integration of the serverless platform with the operating system and CXL/RDMA-based remote memory pools. TrEnv-X ’s core innovations are repurposable sandboxes, which can be shared across different functions to decrease the associated creation overhead, and OS-level memory templates, which enable rapid state restoration from CXL/RDMA-based remote memory pools. To further demonstrate TrEnv-X ’s versatility, we generalize its design from traditional containers for microVM-based agent workloads and introduce new optimizations, including browser sharing and a page cache bypassing mechanism. Our evaluation shows that TrEnv-X achieves up to 7 × reduction in P99 latency and 48% memory savings for container-based functions. When applied to LLM agents, it reduces the P99 latency by up to 58% and memory usage by 61% compared to state-of-the-art systems like E2B.
- 28 March 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Computer Systems
- Vol. 44 (3) , 1-39
Open AccessTrEnv-X: Transparently Share Serverless Execution Environments Across Different Functions and Nodes
Serverless computing is renowned for its computation elasticity, yet its full potential is often constrained by the requirement for functions to operate within local and dedicated background environments, resulting in limited memory elasticity. To address this limitation, this paper introduces TrEnv-X , a co-designed integration of the serverless platform with the operating system and CXL/RDMA-based remote memory pools. TrEnv-X ’s core innovations are repurposable sandboxes, which can be shared across different functions to decrease the associated creation overhead, and OS-level memory templates, which enable rapid state restoration from CXL/RDMA-based remote memory pools. To further demonstrate TrEnv-X ’s versatility, we generalize its design from traditional containers for microVM-based agent workloads and introduce new optimizations, including browser sharing and a page cache bypassing mechanism. Our evaluation shows that TrEnv-X achieves up to 7 × reduction in P99 latency and 48% memory savings for container-based functions. When applied to LLM agents, it reduces the P99 latency by up to 58% and memory usage by 61% compared to state-of-the-art systems like E2B.
- Open Access
- 4 December 2025
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 3 (6) , 1-23
Accelerating Stream Processing Engines via Hardware Offloading
Modern stream processing engines (SPEs) must handle massive real-time data streams under strict latency and throughput requirements. However, conventional SPEs are constrained by their software parallelization strategies (e.g., queue-based data re-partitioning, high synchronization overheads, etc.), which prevent efficient utilization of modern hardware capabilities, ultimately limiting performance scalability. In this paper, we present FlexStream, a novel SPE that leverages hardware offloading to redesign the parallelization strategies and overcome these limitations. By offloading data re-partitioning to hardware and integrating a coupled network-executor model, FlexStream maximizes resource utilization, achieving up to 95% network bandwidth saturation. To address the load imbalance challenges introduced by this design, we implement a lock-free state backend with efficient state migration mechanisms. Overall, FlexStream achieves throughput improvements of 1.95 × - 3.35 × compared to state-of-the-art SPEs (e.g., LightSaber) across six real-world streaming analytics applications. FlexStream cuts latency spikes by 71.9% and migration time by 66.8% during state migration, highlighting the benefits of hardware-software co-design in SPEs. Our work underscores the potential of hardware-software co-design in SPEs, offering a scalable, elastic solution for real-time analytics.
- 4 December 2025
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 3 (6) , 1-23
Open AccessAccelerating Stream Processing Engines via Hardware Offloading
Modern stream processing engines (SPEs) must handle massive real-time data streams under strict latency and throughput requirements. However, conventional SPEs are constrained by their software parallelization strategies (e.g., queue-based data re-partitioning, high synchronization overheads, etc.), which prevent efficient utilization of modern hardware capabilities, ultimately limiting performance scalability. In this paper, we present FlexStream, a novel SPE that leverages hardware offloading to redesign the parallelization strategies and overcome these limitations. By offloading data re-partitioning to hardware and integrating a coupled network-executor model, FlexStream maximizes resource utilization, achieving up to 95% network bandwidth saturation. To address the load imbalance challenges introduced by this design, we implement a lock-free state backend with efficient state migration mechanisms. Overall, FlexStream achieves throughput improvements of 1.95 × - 3.35 × compared to state-of-the-art SPEs (e.g., LightSaber) across six real-world streaming analytics applications. FlexStream cuts latency spikes by 71.9% and migration time by 66.8% during state migration, highlighting the benefits of hardware-software co-design in SPEs. Our work underscores the potential of hardware-software co-design in SPEs, offering a scalable, elastic solution for real-time analytics.
- Top Cited PapersOpen Access
- 15 November 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Storage
- Vol. 22 (4) , 1-38
Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving
Mooncake is the serving platform for Kimi, an LLM chatbot service developed by Moonshot AI. This platform features a KVCache-centric disaggregated architecture that not only separates prefill and decoding clusters but also efficiently utilizes the underexploited CPU, DRAM, SSD and NIC resources of the GPU cluster to establish a disaggregated KVCache. At the core of Mooncake is its KVCache-centric global cache and a scheduler designed to maximize throughput while adhering to stringent latency-related Service Level Objectives (SLOs). Our experiments demonstrate that Mooncake excels in scenarios involving long-context inputs. In tests using real traces, Mooncake increases the effective request capacity by 59% ∼ 498% when compared to baseline methods, all while complying with SLOs. Currently, Mooncake is operational across thousands of nodes, processing over 100 billion tokens daily. In practical deployments, Mooncake ’s innovative architecture enables Kimi to handle 115% and 107% more requests on NVIDIA A800 and H800 clusters, respectively, compared to previous systems.
- 15 November 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Storage
- Vol. 22 (4) , 1-38
Top Cited PapersOpen AccessMooncake: A KVCache-centric Disaggregated Architecture for LLM Serving
Mooncake is the serving platform for Kimi, an LLM chatbot service developed by Moonshot AI. This platform features a KVCache-centric disaggregated architecture that not only separates prefill and decoding clusters but also efficiently utilizes the underexploited CPU, DRAM, SSD and NIC resources of the GPU cluster to establish a disaggregated KVCache. At the core of Mooncake is its KVCache-centric global cache and a scheduler designed to maximize throughput while adhering to stringent latency-related Service Level Objectives (SLOs). Our experiments demonstrate that Mooncake excels in scenarios involving long-context inputs. In tests using real traces, Mooncake increases the effective request capacity by 59% ∼ 498% when compared to baseline methods, all while complying with SLOs. Currently, Mooncake is operational across thousands of nodes, processing over 100 billion tokens daily. In practical deployments, Mooncake ’s innovative architecture enables Kimi to handle 115% and 107% more requests on NVIDIA A800 and H800 clusters, respectively, compared to previous systems.