# 大规模运维 · 日跟踪 · 2026-09-03

> 日期：2026-09-03（Asia/Shanghai）。原料：公开页 + `/workspace/hci-intel/archive/2026-09-03/` 原文。专题正文：`intel/topics/scale-ops.md`。

## 摘要

- **VCF 9.1**：单实例最多 5,000 ESXi 主机（官方称较 9.0 翻倍）；并行升级最多 256 集群；Fleet LCM 声明式。CPU 调度拓扑感知，数据库类数字均为**厂商测试**。
- **Nutanix**：配置上限门户片段每集群 32 主机 / 4,096 开机 VM；NCI 7.5 NVD 称每 PC 10,000 VM / 2,000 主机。Move 6.3 把规模运维延伸到 SP 多租户 Projects。
- **Harvester 1.8**：细粒度 RBAC（Rancher API 强制）、客集群备份、SLE Micro 不可变——补企业控，不补主机上限。
- **Scale Fleet Manager**：ZTP + ALM 应用包舰队分发/回滚/审计；厂商称 1–50,000 站点。
- **k0rdent**：2026-04 自动部署 NVIDIA Run:ai；IREN 2026-08 收购 Mirantis。GPU 调度层被超级控制面吞并。
- **Rafay VMaaS 3.1**（2026-08-27）：NVIDIA 认证双轨道（GB200 + Hopper SXM）+ Token Factory 配额出账。
- **Red Hat**：OpenShift 做 GPU 配额/切分；Day-2 把 Token、检索、工具调用、按团队成本写成运维对象。
- **Proxmox**：无显式节点上限；9.2 新集群默认 token 系数 125 ms；生产有 50+ 节点报告。
- 超聚变「4096 节点」、ZSvirt「10,000 主机」、Scale「50,000 站点」均为厂商/转载口径，**不当保证值**。

## 原文摘录

### [Scale, Simplify, and Secure Your Private Cloud Operations with VCF 9.1](https://blogs.vmware.com/cloud-foundation/2026/05/05/scale-simplify-and-secure-your-private-cloud-operations-with-vcf-9-1/)

> Expanded management scale will allow a single instance to support up to 5,000 ESXi hosts, which will be up 2X from the previous release.
> VCF 9.1 will now offer a 4x increase in parallel upgrade capacity, supporting up to 256 clusters simultaneously.

### [VCF 9.1 Operations What's New](https://techdocs.broadcom.com/us/en/vmware-cis/vcf/vcf-9-0-and-later/9-1/release-notes/vmware-cloud-foundation-9-1-0-0-release-notes/what-s-new/whats-new-vcf-ops.html)

> Manage a maximum of 5000 hosts per VCF Instance
> Per Instance level number of hosts scale support has been increased by 2x from VCF 9.0.
> Parallelization of operations has been enhanced, allowing for 256 simultaneous cluster upgrades and concurrent workflows for domain creation and host commissioning.

### [VCF 9.1 topology-aware CPU scheduler](https://blogs.vmware.com/cloud-foundation/2026/08/27/next-gen-topology-aware-cpu-scheduler-performance-improvements-in-vcf-9-1/)

- 归档：`vmware/next-gen-topology-aware-cpu-scheduler-performance-improvements-in-vcf--dedfbe26.txt`

> Improved performance modeling in VCF 9.1 makes the scheduler topology aware. This means that it considers specific CPU, cache, memory, migration costs, and other factors all based on the platform’s processor architecture.
> SQL Server | Up to 30% higher throughput with HammerDB on vSAN
> Oracle Database | Up to 26% higher throughput in DVD Store

（以上百分比为厂商测试，禁止当招标保证值。）

### [Nutanix NCI 7.5 enterprise-scale management](https://www.nutanix.com/blog/nci-7-5-enterprise-grade-management-built-for-scale-sovereignty-resilience)

> NCI 7.5 delivers validated enterprise-scale architecture (login required) supporting up to 10,000 VMs per Prism Central, 2,000 hosts, and 25,000 categories with 125,000 entity associations.

### [Nutanix Configuration Maximums（访客可见片段）](https://portal.nutanix.com/page/documents/configuration-maximum/list)

> Hosts per cluster | 32 | Maximum number of hosts per cluster
> Powered on VMs per cluster | 4096 | Maximum number of VMs per cluster

### [Move 6.3.0 多租户规模](https://www.nutanix.com/blog/empowering-enterprises-and-service-providers)

- 归档：`nutanix/empowering-enterprises-service-providers-migrations-to-nutanix-project-7e4a2c28.txt`

> To scale efficiently, Service Providers deploy Nutanix Central (NC) to manage all underlying Prism Central (PC) instances. Within Nutanix Central, a logical construct called a Domain establishes a 1:1 mapping with each Prism Central instance.
> A migration plan maps to a single Project Tenant. This automatically restricts VM target selection to Nutanix containers scoped strictly within that tenant's project boundaries.

### [SUSE Virtualization 1.8 RBAC](https://www.suse.com/c/suse-virtualization-1-8-enterprise-controls-migration-composable-platform/)

- 归档：`suse-virtualization/suse-virtualization-1-8-more-control-more-paths-off-vmwa-b456111e.txt`

> Granular RBAC with specific roles. Admins can now assign fine-grained permissions that are API-enforced within Rancher Manager. This means multi-tenant environments get proper access controls without external tooling or manual workarounds.
> Guest cluster backup and restore. Workload protection is now built in at the storage level through the CSI driver. … It’s not a full site-to-site DR solution.
> SLE Micro 6.2 immutable OS. The underlying operating system is now immutable by default. Patches apply as transactional updates that can be rolled back.

### [SC//Fleet Manager ALM](https://www.scalecomputing.com/documents/Data-Sheets/SC-Fleet-Manager-ALM-Data-Sheet.pdf)

> Instead of managing individual deployments, SC//Fleet Manager™ gives users the ability to see and manage their entire fleet at once from an intuitive cloud-based console at fleet.scalecomputing.com.
> 1. Define — Upload an application package … 2. Deploy — Select one or more target clusters and click deploy. 3. Monitor … 4. Update — Publish new versions … with automated validation and rollback controls. 5. Retire — Cleanly remove or replace applications with full audit trails.

### [Red Hat: Managing enterprise AI at scale](https://www.redhat.com/en/blog/managing-enterprise-ai-scale-hosting-deployment-patterns-and-day-2-operations)

- 归档：`redhat/managing-enterprise-ai-at-scale-hosting-deployment-patterns-and-day-2--48ec8498.txt`

> Red Hat OpenShift solves operational problems that are hard to get right when self-hosting models: GPU quotas and isolation across teams, hybrid accelerator support, and AI components as standard container workloads. That includes sharing individual GPUs through partitioning or time-slicing.
> Day 2 operations are what keep a production AI system stable after it goes live. … Peak concurrency and token throughput matter more than average load.
> Cost broken down by application, team, and model makes the financial and capacity implications of that traffic visible.

### [Mirantis k0rdent × NVIDIA Run:ai](https://www.mirantis.com/company/press-center/company-news/mirantis-automates-ai-factory-deployments-with-k0rdent-ai-and-nvidia-run-ai/)

> CAMPBELL, Calif., April 15, 2026 – Mirantis … announced the company is enabling enterprises and neocloud providers to deploy production-ready AI platforms in minutes … by automated deployment and lifecycle management of NVIDIA Run:ai through the Mirantis k0rdent AI platform.
> Mirantis executed more than 100 functional tests covering workload submission, scheduling behavior, multi-tenant operations, and platform lifecycle management, achieving partner-certified status.

### [Proxmox Cluster Manager](https://pve.proxmox.com/pve-docs/chapter-pvecm.html)

> There’s no explicit limit for the number of nodes in a cluster. In practice, the actual possible node count may be limited by the host and network performance. Currently (2021), there are reports of clusters (using high-end enterprise hardware) with over 50 nodes in production.
> The Proxmox VE cluster stack requires a reliable network with latencies under 5 milliseconds (LAN performance) between all nodes to operate stably.
> Since Proxmox VE 9.2, new clusters are created with a lower token coefficient of 125 milliseconds explicitly set in /etc/pve/corosync.conf.

### [William Lam: VCF Fleet 没有固定 Instance 个数](https://williamlam.com/2025/10/how-many-vmware-cloud-foundation-vcf-instances-can-a-vcf-fleet-support.html)

> Unlike traditional configuration maximums which only looks at the number of vCenter Servers, vSphere Clusters and Hosts, VCF Operations uses a different measurement, which is expressed as Objects and Metrics.
> Extra Large deployment size of VCF Operations … 1 x Node VCF Operations: Up to 100K objects and 20M metrics; 16 x Node VCF Operations: Up to 1M objects and 126M metrics.

## 对专题档案的影响

- 保证值只采用 **VCF 发行说明 5,000/256** 与 **Nutanix 配置上限门户 32/4096**；NVD 10,000 VM 需登录文档，正文已标。
- AI 运维对象从主机变成 **GPU 配额 + Token + 拓扑 + 故障域**；Rafay/k0rdent/OpenShift AI 进智算短名单，不进 VMware 替换短名单。
- 深信服 vkey/指纹主机、Vates 3 主机入门墙、OVE 容器墙、Nutanix Ultimate 墙，全部是「规模运维第二张账单」。