Challenges and Emerging Trends in Cloud and Internet Storage Architecture

The storage domain of cloud and internet services currently relies on server‑deployed distributed storage, facing four major challenges: (1) data lifecycle periods that do not match server update cycles, (2) difficulty achieving both high performance reliability and high resource utilization, (3) the demand for lightweight, high‑bandwidth, low‑latency shared storage from serverless and AI/ML workloads, and (4) the so‑called “data‑center tax” where CPU‑centric architectures waste up to 30% of compute power on I/O.

To address these issues, recent hardware trends provide a foundation for a new storage‑compute disaggregation architecture. High‑performance Ethernet‑Bunch‑of‑Flash (EBOF) and similar disk‑less chassis replace traditional server disks, while the NVMe‑over‑Fabric (NoF) protocol, CXL, and other high‑speed interconnects enable flexible, low‑latency communication between processors, memory, and storage. Specialized data‑processing units (DPUs, IPUs) and programmable network devices further offload storage‑related tasks, improving energy efficiency.

The proposed next‑generation disaggregated architecture features three key characteristics: (1) a thorough decoupling of compute and storage resources into independent pools (CPU, memory, HDD/SSD), allowing independent scaling and sharing; (2) fine‑grained task partitioning where DPUs and other accelerators replace general‑purpose CPUs for data‑intensive operations; and (3) a unified resource pool that can be dynamically composed to meet diverse workload requirements.

Four architectural modules are defined: Storage Module – employing EBOF, Ethernet‑Bunch‑of‑Memory (EBOM), and Ethernet‑Bunch‑of‑Disk (EBOD) chassis with NoF networking to provide block and file services; Compute Module – leveraging DPUs and other specialized processors to form power‑efficient compute pools; High‑Throughput Data Bus – using CXL, NoF, and IP‑based protocols to achieve sub‑microsecond latency and support both hot and cold data paths; and Network Fabric – integrating 10‑GbE, RDMA, NVMe/RoCE, and emerging memory‑centric fabrics to enable flexible, high‑bandwidth connectivity.

By reorganizing storage, networking, and compute resources into these modular pools, the architecture aims to reduce storage waste, lower latency, improve energy efficiency, and better serve emerging cloud and internet workloads such as serverless applications, AI/ML, and large‑scale data analytics.