# 存储专题日跟踪 2026-09-03

## 摘要

**今日无新的 HCI 存储版本号或报价页。** 最强信号是 **2026-09-01** MLCommons 发布的 MLPerf Storage **v3.0**：正式加入 KV Cache、向量库，并加 S3 访问层（本轮约 1/6 提交走 S3）。checkpoint 出现 877 GiB/s 量级的厂商提交；HCI 软件基本缺席。这把「虚机盘」和「AI 数据面」的产品分界写进了可引用的行业测试。次强信号仍是 8 月 Gartner 换平台窗口，以及 Nutanix 把外置阵列路径写成 **FC → NVMe/TCP**。vSAN ESA 的私有 AI 存储设计（100 GbE + 可选 RoCE）是 VMware 侧仍有效的对照页。

规划即时含义：本周话术同时带 **dHCI/NVMe/TCP 保阵列** 和 **不要用 HCI 数据盘扛 checkpoint/KV**。三平面 QoS 比再报一个 IOPS 更有用。

## 原文摘录

### 1. MLPerf Storage v3.0（2026-09-01）——今日仍是最强存储新闻

- URL：https://mlcommons.org/2026/09/mlperf-storage-v3-0-results/

> SAN FRANCISCO, Sept. 01, 2026 — Today, MLCommons® announced the results of its industry-standard MLPerf® Storage v3.0 benchmark suite
> Version 3.0 adds a new KV Cache test, which measures storage performance for LLM inference cache read/write operations.
> It also adds a new Vector Database (VDB) test that measures storage performance for vector indexing and querying workloads.
> In addition, Version 3.0 adds new support for an S3 object storage access layer, alongside and as an alternative to the existing POSIX-compliant layer. … Approximately one-sixth of the total submissions in this round utilized the S3 storage access layer.
> “As the scale of AI contexts reaches into the trillions, we expect object-based storage systems to emerge as a viable – and possibly preferred – alternative to filesystem-based storage,” said Curtis Anderson, MLPerf Storage working group co-chair.
> Nineteen organizations submitted to this round … eleven first-time submitters: Azure, Everpure, HolmesAI, Nebius, NewFW, NVIDIA, OpenLake, Suzhou Zishan Longlin, TuringData, XSKY, and ZettaLane

### 2. 877 GiB/s checkpoint 与缺席名单（行业媒体，同日轮次）

- URL：https://www.storagereview.com/news/mlperf-storage-v3-0-877-gib-s-checkpoints-a-cloud-first-and-a-leaderboard-turned-over

> In the 1.25T parameter checkpointing simulation with 1,024 client accelerators, the 30-node configuration sustained 877.52 GiB/s of write bandwidth over a 17.74-second operation, alongside 588.28 GiB/s of read bandwidth completed in 28.99 seconds.
> Just as notable is that DDN, Huawei, Hammerspace, and Lightbits, the names that defined the v2.0 leaderboard, did not submit this round.

### 3. Gartner：hypervisor 换平台正在改存储采购（2026-08-27 转述 8-19 MQ）

- URL：https://virtualizationreview.com/articles/2026/08/27/gartner-hypervisor-replatforming-is-reshaping-enterprise-storage.aspx

> The Aug. 19 2026 Magic Quadrant for Enterprise Storage Platforms says widespread licensing restructuring and pricing changes are prompting organizations to reconsider hypervisor strategies, creating what Gartner calls a "major displacement window." That transition is changing storage buying decisions along with virtualization decisions.

### 4. Nutanix：外置存储从 FC 迁到 NVMe/TCP

- URL：https://www.nutanix.com/tech-center/blog/nutanix-cloud-platform-with-external-storage

> Organizations utilizing Fibre Channel (FC) as a storage access protocol to connect to external storage can effortlessly convert their existing infrastructure to NVMe/TCP upon adopting the Nutanix Cloud Platform (NCP) with external storage solution.
> Just as iSCSI democratized SAN storage by leveraging Ethernet, NVMe/TCP extends the NVMe protocol across standard Ethernet networks using the TCP transport binding defined in the NVMe TCP Transport Specification.
> A FlashArray can run Fibre Channel and NVMe/TCP concurrently, so both protocols can be active at the same time.

### 5. vSAN ESA 私有 AI 存储设计（文档 Last Updated January 7, 2026）

- URL：https://techdocs.broadcom.com/us/en/vmware-cis/vcf/vvs/1-0/private-ai-ready-infrastructure-for-vmware-cloud-foundation/detailed-design-for-ai-ready-infrastructure/storage-design.html

> vSAN ESA (Express Storage Architecture) is optimized for performance … when deployed with 100 Gbps networking and faster. vSAN ESA also supports RDMA over Converged Ethernet (RoCE v2).
> You can store such vector databases on vSAN clusters in a VI workload domain, mitigating noisy neighbor events and resource contention with productive inference workloads.
> External storage solutions can be leveraged as supplemental storage … S3-compatible object storage and NFS exports, provided directly to containers and virtual machines together with vSAN for shared datasets, model repositories, and archival purposes.
> AIR-STORAGE-001 Use vSAN ESA with 100 Gbps networking and,if possible, RDMA. … the minimum bandwidth for vSAN ESA is 25 Gbps

### 6. 深信服 aSAN（产品页，非今日稿；指标为厂商自称）

- URL：https://www.sangfor.com.cn/sangfor-cloud/asan

> aSAN 单物理节点可提供超20 万IOPS，能够为Oracle RAC、SQL Server Alway on 等核心数据库提供优异的性能体验
> aSAN 与服务器虚拟化构建的超融合基础架构提供延伸集群存储双活方案，实现RPO=0，RT0 秒级的机房故障恢复。

### 7. KV-cache 基准在测什么（MLCommons 仓库说明）

- URL：https://github.com/mlcommons/storage/tree/main/storage/kv_cache_benchmark

> During LLM inference, models store intermediate attention data in a structure called the KV (Key-Value) cache. This cache grows with conversation length and can consume enormous amounts of memory. Production systems offload this cache from expensive GPU VRAM to cheaper CPU RAM or NVMe storage.

## 出处

- https://mlcommons.org/2026/09/mlperf-storage-v3-0-results/
- https://www.storagereview.com/news/mlperf-storage-v3-0-877-gib-s-checkpoints-a-cloud-first-and-a-leaderboard-turned-over
- https://www.hpcwire.com/2026/09/01/mlcommons-bolsters-mlperf-storage-benchmark-with-kv-cache-vector-db-and-s3-support/
- https://virtualizationreview.com/articles/2026/08/27/gartner-hypervisor-replatforming-is-reshaping-enterprise-storage.aspx
- https://www.nutanix.com/tech-center/blog/nutanix-cloud-platform-with-external-storage
- https://www.nutanix.com/press-releases/2026/nutanix-and-netapp-form-strategic-alliance
- https://techdocs.broadcom.com/us/en/vmware-cis/vcf/vvs/1-0/private-ai-ready-infrastructure-for-vmware-cloud-foundation/detailed-design-for-ai-ready-infrastructure/storage-design.html
- https://www.sangfor.com.cn/sangfor-cloud/asan
- https://developer.nvidia.com/blog/nvidia-bluefield-4-powers-new-scale-in-network-infrastructure-for-agentic-ai-factories/