Why Traditional Distributed Storage Struggles and How New Compute‑Storage Separation Can Transform Cloud Data Centers

Challenges in Current Distributed Storage

From cloud and internet business scenarios, storage domains rely on server‑deployed distributed storage services, facing several challenges: (1) mismatch between data retention cycles and server update cycles, requiring long‑term data to be kept according to lifecycle policies (typically 8‑10 years); (2) difficulty balancing performance, reliability and resource utilization—performance‑oriented storage often uses three‑replica or two‑replica RAID, achieving only about 30 % space utilization; (3) emerging serverless applications demand lightweight, high‑bandwidth, low‑latency shared storage; (4) the “datacenter tax” where CPU‑centric architectures waste up to 30 % of compute power on I/O handling and exhibit low energy‑efficiency.

Hardware Technology Trends

Recent advances such as Ethernet Bunch of Flash (EBOF), NoF (NVMe over Fabric), high‑performance disk frames (e.g., OpenFlex, Ceres), and the emergence of CXL, DPU, IPU, and programmable switches are reshaping storage and compute. CXL, standardized in 2016, enables fast interconnect between CPUs, GPUs, FPGAs, and accelerators. Industry also explores remote memory pooling, erasure‑code based storage, and new network standards like Gen‑Z and OpenCAP.

Features of the New Compute‑Storage Separation Architecture

The new architecture emphasizes (1) complete decoupling of compute and storage resources into independent pools (CPU, memory, HDD/SSD); (2) fine‑grained task division where specialized accelerators (DPUs, data processing units) handle workloads unsuitable for general‑purpose CPUs, improving energy efficiency.

Key characteristics include diskless servers with remote storage pools, diverse network protocols (CXL+NoF+IP) achieving sub‑microsecond latency, specialized data processors offloading erasure‑code, encryption, compression, and extremely high storage density through disaggregated storage systems that combine traditional RAID with modern high‑density disk frames.

Three‑Tier Design for Cloud and Internet Scenarios

The architecture reorganizes resources into three simplified layers: storage modules (EBOF, Ethernet Bunch of Memory, EBOD, etc.) providing block and file services via NoF; a high‑throughput bus network (CXL, RDMA, RoCE) enabling fast memory pooling; and compute modules built from DPUs and other accelerators to deliver optimal performance‑per‑watt.

Target Application Scenarios

Three typical workloads benefit: (1) virtualization where local disks are pulled away and replaced by remote disaggregated storage; (2) hot‑data services such as databases and big‑data platforms requiring large memory and key‑value interfaces; (3) containerized workloads (Ceph, Lustre) that offload data layout and enable tiered storage from hot to cold media, improving overall resource utilization.