Hong Jiang Ocean University Of China Team

Introduction

The Hong Jiang Ocean University of China team refers to the distinguished research group led by Professor Hong Jiang (江鸿) at the Ocean University of China (OUC), located in Qingdao, Shandong Province. Their work bridges the gap between theoretical computer science and practical systems engineering, addressing the critical I/O bottlenecks that plague modern data-intensive scientific discovery and enterprise applications. Plus, as a central force within the School of Computer Science and Technology (formerly the College of Information Science and Engineering), this team has established a formidable reputation in the global computer science community, particularly in the domains of high-performance computing (HPC), parallel storage systems, big data processing, and non-volatile memory (NVM) architectures. Professor Jiang, an IEEE Fellow and a highly cited researcher, directs a group that consistently publishes in top-tier venues such as IEEE Transactions on Computers, ACM Transactions on Storage, FAST, SC, and ICS. Understanding the contributions of this team offers valuable insight into the cutting edge of storage hierarchy optimization and the future of exascale computing infrastructure That's the whole idea..

Short version: it depends. Long version — keep reading.

Detailed Explanation

The Leadership and Vision of Professor Hong Jiang

Professor Hong Jiang serves as the intellectual anchor and strategic visionary of the team. With a Ph.D. Plus, in Computer Science from the University of Nebraska-Lincoln and a career spanning both academia and industry research labs (including a tenure at the University of Texas at Arlington), he brings a unique blend of theoretical rigor and systems pragmatism to OUC. His research philosophy centers on the concept of "data-centric computing," arguing that as compute capabilities (FLOPS) have grown exponentially via Moore’s Law and specialized accelerators (GPUs, TPUs), the movement of data has become the primary bottleneck. The team’s mission, therefore, is to re-architect the storage stack—from the physical media (SSD, NVM, HDD) through the file system and middleware layers up to the application interface—to minimize data movement, maximize throughput, and ensure data integrity and security. Under his guidance, the group has secured substantial funding from the National Natural Science Foundation of China (NSFC), the National Key R&D Program of China, and major industry partnerships with companies like Huawei, Inspur, and Sugon The details matter here..

Organizational Structure and Research Culture

The Hong Jiang team at OUC typically comprises 15–25 members, including tenured associate professors, postdoctoral researchers, Ph.Think about it: this commitment to "systems that work" has resulted in several open-source releases (e. So d. Think about it: the group operates with a "project-based" sub-structure, where smaller squads of 3–5 students tackle specific sub-problems—such as NVMe-over-Fabrics optimization, erasure coding for distributed storage, or AI-driven prefetching—while maintaining weekly group seminars for cross-pollination of ideas. g.candidates, and master’s students. Unlike groups that focus solely on simulation, the Hong Jiang team builds real prototypes on hardware testbeds featuring the latest Intel Optane DC Persistent Memory, Samsung Z-NAND SSDs, and high-speed RDMA networks (100GbE/200GbE InfiniBand). A hallmark of their culture is the emphasis on artifact evaluation and open-source contribution. , optimized key-value stores, log-structured file system prototypes) that are actively used by the broader research community Simple as that..

Step-by-Step Concept Breakdown: Core Research Thrusts

The team’s research portfolio can be deconstructed into four interconnected pillars. Each pillar addresses a specific layer of the storage hierarchy, and breakthroughs in one often catalyze advances in the others.

1. Next-Generation Storage Media Management (The Device Layer)

The advent of Non-Volatile Memory (NVM), such as Intel 3D XPoint (Optane), and advanced Flash technologies (QLC NAND, ZNS SSDs) has disrupted the traditional block-interface paradigm. The team’s work here focuses on:

• Byte-Addressable Persistence: Designing file systems (e.g., NOVA-Fortis variants) and key-value stores (e.g., PMEM-KV optimizations) that put to work load/store instructions directly on persistent media, bypassing the heavy kernel block layer.
• Zoned Namespace (ZNS) & Open-Channel SSDs: Developing zone-aware garbage collection and write placement algorithms that exploit the sequential-write-only constraint of ZNS drives to eliminate write amplification and reduce tail latency.
• Cross-Layer Resilience: Creating application-consistent crash recovery mechanisms that put to use hardware atomic write units (e.g., 4KB/8KB atomicity on Optane) to avoid expensive journaling overhead.

2. High-Performance Parallel File Systems & Distributed Storage (The Cluster Layer)

For HPC clusters and cloud infrastructure, the team optimizes the distributed file system stack (Lustre, Ceph, BeeGFS, and proprietary systems like OceanFS).

• Metadata Scalability: They have proposed distributed metadata indexing using learned indexes (B-tree/ML hybrids) and RDMA-based metadata caching to scale metadata operations to billions of files—a critical requirement for AI/ML training datasets.
• Burst Buffer Architecture: The team designs multi-tier burst buffer systems (NVM + SSD + HDD) with data-aware tiering policies. Their algorithms predict I/O access patterns (sequential vs. random, read vs. write) using lightweight online learning to stage data optimally, reducing checkpoint/restart times for scientific simulations by 40–60%.
• Erasure Coding Optimization: They developed pipeline-friendly, locality-aware erasure codes (e.g., Clay codes, Pipeline codes) that drastically reduce network traffic during degraded reads and reconstruction, a major pain point in geo-distributed storage.

3. Big Data & AI I/O Acceleration (The Application Layer)

Recognizing that modern workloads (TensorFlow, PyTorch, Spark, Flink) have vastly different I/O signatures than traditional HPC (MPI-IO), the team builds domain-specific storage engines.

• Deep Learning I/O: They created data loading accelerators that fuse shuffling, augmentation, and prefetching into the storage layer, utilizing GPUDirect Storage (GDS) to DMA data directly from NVMe/NVM to GPU memory, bypassing CPU bounce buffers.
• Graph Processing: For iterative graph analytics (PageRank, GNNs), they designed **vertex-centric persistent

storage engines that make use of asynchronous I/O (io_uring) to mask the latency of random access patterns. * Stream Processing & Log-Structured Engines: To support real-time analytics in Apache Flink and Kafka, the team implements zero-copy log-structured merge (LSM) trees optimized for high-throughput ingestion. On the flip side, by decoupling the compute-heavy graph traversal from the I/O-heavy vertex updates, they achieve near-linear scaling on massive, irregular datasets. These engines minimize write amplification by aligning internal compaction cycles with the physical NAND erase blocks of the underlying storage hardware.

4. Emerging Frontiers: Computational Storage & Intelligent Fabric

The final frontier of the team's research lies in moving the compute closer to the data to break the "Von Neumann bottleneck."

• Computational Storage Drives (CSDs): They are pioneering in-situ data processing where simple filtering, aggregation, and compression operations are offloaded to the SSD controller. This reduces the volume of data transferred over the PCIe bus, effectively turning the storage device into a co-processor.
• SmartNIC-Integrated Storage: By leveraging DPU (Data Processing Unit) technology, the team implements offloaded storage virtualization. This allows for line-rate encryption, compression, and deduplication to be performed within the network fabric, ensuring that CPU cycles are reserved exclusively for application logic.

Conclusion

The evolution of storage technology is no longer merely a race for higher capacity or raw throughput; it is a multidimensional challenge involving hardware-software co-design, intelligent orchestration, and domain-specific optimization. By bridging the gap between the physical properties of non-volatile media and the complex requirements of AI, HPC, and Big Data workloads, the research presented here moves us toward a future of "transparent performance." In this future, the storage subsystem is no longer a passive repository, but an active, intelligent, and highly integrated component of the global computing fabric Most people skip this — try not to..

5. Edge‑to‑Cloud Continuity and Multi‑Tiered Analytics Fazer

While high‑performance datacenters provide the raw horsepower for training, the majority of inference traffic is now shifting to the edge—smartphones, autonomous vehicles, and industrial IoT gateways. The research team has begun to bridge this divide by designing heterogeneous storage fabrics that span cloud, edge, and fog nodes.

• Adaptive Tiering with Predictive Workload Migration – By continuously profiling access patterns and embedding lightweight machine‑learning models on the controller, the system can pre‑emptively migrate hot data from NVMe SSDs to faster, but more expensive, 3D‑XPoint or even persistent memory (PMEM) tiers. This reduces inference latency for latency‑sensitive workloads without incurring the high cost of keeping everything in the fastest tier That alone is useful..

• Federated Storage Orchestration – Leveraging a lightweight, protocol‑agnostic API, the team orchestrates data placement across geographically dispersed nodes. Federated queries can be executed in a single pass, with the engine automatically pulling the smallest representation of the data from the nearest edge cache, thereby minimizing WAN traffic No workaround needed..

• Hybrid Cloud Bursting – During peak training phases, the system can naturally offload excess storage and compute to public cloud backends. The same data‑centric APIs used in the on‑prem cluster are exposed on the cloud, ensuring that data locality decisions remain consistent across environments.

6. Security, Privacy, and Trust in Data‑Intensive Systems

As data traverses multiple tiers and jurisdictions, the need for dependable security mechanisms becomes key. The research team has integrated several novel safeguards directly into the storage stack.

• Hardware‑Based Zero‑Trust Encryption – Each storage node embeds an SGX‑compatible enclave that performs encryption/decryption on COUNTER‑mode streams. Keys are derived from a hardware root of trust and never leave the enclave, eliminating the attack surface presented by software key managers.

• Differential Privacy at the Storage Layer – By attaching a privacy‑budget manager to the log‑structured merge engine, the system can inject calibrated noise into query results before they leave the storage node. This ensures that even if an adversary intercepts the data stream, the privacy guarantees hold.

• Tamper‑evident Logging – Every write operation is hashed and appended to a Merkle tree stored in a secure element. The resulting root hash is periodically signed and pushed to a distributed ledger, providing an immutable audit trail that can be verified by stakeholders worldwide.

7. Toward Neuromorphic and Quantum‑Enabled Storage

Looking further ahead, the team is exploring storage paradigms that emulate biological brains and quantum mechanics Not complicated — just consistent. Nothing fancy..

• Spike‑Based Non‑Volatile Memory (SNVM) – By combining memristive crossbar arrays with event‑driven controllers, the research group has prototyped a storage fabric that operates on spikes rather than voltage levels. This allows for ultra‑low‑power,-tailored data retention suitable for continuousmoisture sensors and brain‑computer interfaces That alone is useful..

• Quantum‑Backed Key Management – Leveraging quantum key distribution (QKD) channels, the system can secure encryption keys with unconditional security. Even if a classical adversary gains access to the storage hardware, the QKD‑derived keys remain mathematically unbreakable.

• Hybrid Quantum‑Classical Accelerators – For specific workloads such as factorization or lattice simulations, the storage engine can hand off data to a quantum accelerator via a dedicated Q‑PCIe link, thus ensuring that the data residency and movement overheads are minimized And that's really what it comes down to. Took long enough..

Conclusion

The convergence of high‑throughput, low‑latency storage with intelligent, domain‑specific optimizations is redefining what it means to “store” data in modern computing. By fusing advanced hardware—NVMe‑o, 3D‑XPoint, persistent memory—with software abstractions that expose data locality, consistency, and security as first‑class citizens, the research team has laid out a roadmap toward storage systems that are not merely passive repositories but active participants in the computation pipeline.

From data‑loading accelerators that eliminate CPU bottlenecks, to vertex‑centric engines that tame irregular graph workloads, to zero‑copy log‑structured trees that honor the physical realities of NAND flash, each innovation addresses a concrete pain point in today’s AI, HPC, and big‑data ecosystems. Coupled with emerging frontiers such as computational storage, smart‑NIC virtualization, and neuromorphic memory, the future promises a truly integrated, end‑to‑end fabric where data movement, processing, and protection are co‑designed from the ground up.

In this paradigm, storage is no longer a passive “

active substrate—it becomes an intelligent, programmable layer that understands the semantics of the data it holds, anticipates access patterns, and participates in the very algorithms that drive insight. The boundaries between memory, compute, and network continue to dissolve, giving rise to a continuum where data gravity is managed not by moving bytes to processors, but by moving processing to the bytes, wherever they reside Worth keeping that in mind. Which is the point..

This shift demands a corresponding evolution in system software, programming models, and operational practices. Practically speaking, developers must think in terms of data‑centric workflows, leveraging APIs that express intent—“filter this column,” “traverse this subgraph,” “verify this ledger”—rather than low‑level block commands. Runtime systems will negotiate placement, replication, and acceleration transparently, guided by real‑time telemetry and policy engines that balance latency, energy, and regulatory constraints Worth knowing..

The bottom line: the research agenda outlined here points toward a self‑optimizing storage fabric that learns from its workloads, heals from failures, and secures itself against both classical and quantum threats. As the volume and velocity of data grow beyond the reach of traditional architectures, such a fabric will be the foundation upon which the next generation of scientific discovery, real‑time AI, and trustworthy digital infrastructure is built. The journey from passive repository to active participant is well underway; the remaining challenge is to make this intelligence ubiquitous, interoperable, and accessible to every layer of the computing stack.