Recently, four papers by faculty member Yangyang Wang of our School were accepted at the prestigious international conferences INFOCOM 2026, ICDE 2026, DASFAA 2026, and IWQoS 2026. The papers focus primarily on distributed consensus protocols and distributed systems, with Nanchang University as the sole institutional affiliation. These research outcomes provide theoretical foundations and technical insights for the efficient processing of big data.

1.Paper Title: CD-Raft: Reducing the Latency of Distributed Consensus in Cross-Domain Sites

Authors: Yangyang Wang, Ziqian Cheng, Yucong Dong and Zichen Xu*

Conference: INFOCOM 2026, CCF Class A

Paper Overview: As data centers are increasingly deployed across multiple geographical regions and administrative domains, the demand for cross-domain data processing and disaster recovery continues to grow. Traditional leader-based consensus protocols often incur long cross-domain round-trip times (RTTs) when handling cross-domain requests, significantly increasing system response latency. To address this issue, we propose CD-Raft, an optimized Raft protocol for cross-domain environments. CD-Raft introduces two types of roles, Domain Leader and Global Leader, and incorporates two core mechanisms: Fast Return and Optimal Global Leader Election. Fast Return allows the Domain Leader in the client’s local domain to respond once the required safety conditions are satisfied, reducing the request-response path to approximately one cross-domain RTT. Optimal Global Leader Election selects the location of the Global Leader according to the request distribution across domains and inter-domain latency, thereby reducing overall system latency. We implemented a CD-Raft prototype in Go and built its storage and communication modules using RocksDB, gRPC, and Protocol Buffers. We then conducted YCSB experiments in a real cross-domain network environment on Huawei Cloud. Under a representative balanced read-write workload, CD-Raft reduced average latency by 32.90% and 99th-percentile tail latency by 49.24% compared with Raft. These results demonstrate that CD-Raft can effectively reduce request-processing latency for cross-domain consensus protocols and improve overall system performance in cross-domain scenarios.

2.Paper Title: Nezha: A Key-Value Separated Distributed Store with Optimized Raft Integration

Authors: Yangyang Wang, Yucong Dong, Ziqian Cheng and Zichen Xu*

Conference: ICDE 2026, CCF Class A

Paper Overview: With the rapid development of big-data applications, cloud-native services, and distributed database systems, distributed key-value stores have become essential infrastructure for modern data-intensive applications. To ensure data consistency in multi-replica environments, existing systems commonly employ consensus protocols such as Raft together with LSM-tree storage engines to provide reliable data persistence. However, our systematic analysis shows that in conventional Raft-based distributed key-value stores, a single write request often incurs multiple disk writes, including Raft log persistence, storage-engine WAL writes, and MemTable flushes. This creates substantial redundant persistence overhead between the consensus and storage layers and becomes a major bottleneck for system throughput and latency. To address this issue, we propose Nezha, a key-value-separated distributed store that tightly integrates key-value separation with the Raft consensus protocol, significantly reducing write amplification and redundant I/O while preserving strong consistency. Specifically, Nezha introduces KVS-Raft, a protocol optimized for key-value storage. Value data are persisted once in the Raft ValueLog, while the state machine maintains only lightweight offset information, reducing the repeated persistence of values in conventional systems to a single write. To address the garbage-collection overhead and read-performance degradation introduced by integrating key-value separation with Raft, Nezha further designs a Raft-aware garbage-collection mechanism and a three-stage request-processing workflow. Different data-access strategies are adopted before, during, and after GC, ensuring that Put, Get, and Scan requests are processed correctly while the storage layout is optimized. Experimental results show that, compared with conventional Raft-based key-value stores, Nezha improves average throughput for Put, Get, and Scan operations by 460.2%, 12.5%, and 72.6%, respectively. These results demonstrate the effectiveness of jointly optimizing the consensus and storage layers in high-performance distributed storage systems.

3.Paper Title: P-Raft: Distributed Consensus with Predictive Optimization Under Cross-Domain Sites

Authors: Yangyang Wang, Ziqian Cheng, Yucheng Ji and Zichen Xu*

Conference: DASFAA 2026, CCF Class B

Paper Overview: As data centers continue to expand across multiple geographical regions and areas, cross-domain data processing has become a common scenario in distributed databases. Traditional leader-based consensus protocols typically need to wait for cross-domain replica propagation and acknowledgments when processing cross-domain requests, causing the request-response path to be significantly affected by inter-domain communication latency. Meanwhile, network conditions and request workloads change dynamically, making fixed or reactive leader-selection strategies unable to adapt promptly to workload hot-spot migration. To address these issues, we propose P-Raft, a Raft variant designed for cross-domain environments. P-Raft combines Fast Commit, an optimal leader-domain evaluation model, and a prediction-based leader-migration mechanism. Fast Commit confines the commit path to the Leader’s domain, reducing the impact of cross-domain communication on the response path. The evaluation model jointly considers inter-domain latency and the read-write workloads of each domain to identify a better Leader location. The prediction mechanism uses machine learning to forecast future workload changes and migrate the Leader in advance. We implemented a P-Raft prototype based on etcd and evaluated it using YCSB. Under a balanced read-write workload, P-Raft reduced average latency by 75.19%, 72.61%, and 63.89% compared with Raft, EPaxos, and GeoLM, respectively. These results show that P-Raft can effectively reduce access latency in cross-domain consensus systems and improve system performance under dynamic workloads.

4.Paper Title: Semantic Raft: A Fault-Tolerant Multi-Agent Inference Framework for Reliable LLM Services

Authors: Yidong Su, Yangyang Wang*, Yucong Dong and Zichen Xu

Conference: IWQoS 2026, CCF Class B

Paper Overview: As large language models are increasingly applied to critical scenarios such as financial auditing, code generation, and complex logical reasoning, their reliability in long-horizon reasoning has become an increasingly prominent concern. Existing single-model chain-of-thought reasoning methods are vulnerable to early hallucinations, causing erroneous intermediate states to accumulate and amplify throughout subsequent reasoning. Although multi-agent debate methods introduce redundant reasoning, the absence of strict state-consistency constraints makes agents prone to blind agreement and correlated failures. To address these problems, we propose Semantic Raft, a fault-tolerant multi-agent inference framework for large language model services. Inspired by the Raft consensus protocol in distributed systems, Semantic Raft models multi-step reasoning as a semantically replicated state machine and constructs a logical firewall at every reasoning step through reputation-weighted Leader election, atomic reasoning proposals, prompt-driven blind verification, and majority-consensus commit. Specifically, the system first selects a Leader based on the agents’ historical reliability and their confidence in the current task. The Leader generates a candidate reasoning step, after which the other agents reason independently under logical isolation and verify whether the proposed step is valid. Only results approved by a majority are appended to the global reasoning log. Through this structured “propose-verify-commit” protocol, Semantic Raft effectively blocks the propagation of erroneous states and mitigates hallucination, blind conformity, and cascading failures. Experimental results show that the framework significantly outperforms existing baselines on mathematical reasoning tasks under high agent failure rates and on long-horizon logical reasoning tasks, providing a new systematic approach to building reliable large language model services.

The acceptance of these four papers marks a significant breakthrough for our School in the field of big-data research. Going forward, the School will continue to encourage faculty and students to pursue research at the international frontier, deepen their work on foundational methods, and broaden interdisciplinary applications. It will strive to produce more high-quality research outcomes in big data and artificial intelligence, further strengthen the discipline’s research foundation, and continuously enhance the School’s academic influence and international competitiveness.

Reviewed by: Zeng Qingwei, Xu Zichen,  and Gan Lijiao