Синькуань Сервер | Поставщик оборудования для корпоративных серверов
Scale-Up vs Scale-Out：SAP HANA 部署的第一个决策
SAP HANA 的硬件选型始于一个架构决策，而不是 CPU 型号：你的工作负载应该跑在一个大内存的 4 路节点上，还是拆成多个小节点？两种路线各有适用场景：
| Измерение | Scale-up (4-розетка) | Scale-out (multi-node) |
|---|---|---|
| Node count | 1 (начальный) | 2-10+ |
| Memory ceiling | Вплоть до 16 TB per node | Aggregated, 1-8 TB per node |
| Data distribution | Single instance, no partitioning | Partitioned across nodes |
| Complexity | Low — one system to manage | Higher — distribution, сеть, licensing |
| Best fit | <16 TB workloads, ОЛТП, консолидация | Very large analytics, мультиарендатор, extreme scale |
The rule of thumb in 2026: if your HANA working set fits under ~16 TB, scale-up wins on total cost and operational simplicity — one database, no partitioning, no cross-node networking. That is precisely the niche the 4-socket PowerEdge R960 occupies.
SAP HANA’s Hardware Iron Rules
SAP certifies HANA on a strict hardware recipe. The non-negotiables:
- Memory dominates everything. HANA keeps the working set in RAM — capacity and bandwidth beat core count. 64 DDR5 slots at 4800 MT/s is the scale-up sweet spot.
- Fast durable storage for the log. The redo log must persist with sub-millisecond latency — NVMe, not spinning disk.
- High-bandwidth client network. 10/25GbE minimum; 100GbE for consolidation.
- Fast restart. After a crash, HANA reloads from data volumes — NVMe cuts recovery windows from hours to minutes.
The PowerEdge R960 as a Scale-Up Foundation
Dell positions the R960 precisely for databases and business-critical applications. The hardware map for a HANA-certified build:
- 4 x 4th Gen Intel Xeon Scalable (60 ядра каждое) — SAP-certified CPU generations with Intel QuickAssist optional.
- 64 x DDR5 RDIMM slots, вплоть до 16 туберкулез в 4800 МТ/с — the memory ceiling that keeps scale-up viable.
- Вплоть до 24 x NVMe U.2 (368.64 ТБ) for data volumes, или 32 x SAS/SATA for tiered layouts.
- 12 x PCIe Gen5 slots — HCA, dual 100GbE adapters and storage HBAs coexist without contention.
- BOSS-N1 M.2 hardware-RAID boot — OS boots off dedicated modules, keeping all drive bays for data.
- iDRAC9 + OpenManage Enterprise — out-of-band control, telemetry and vCenter integration for the operations team.
For scale-out estates that genuinely outgrow 16 ТБ, the xFusion 2288H V7 (2-розетка, вплоть до 28 NVMe) makes an efficient worker node — see our server insights for the platform landscape.
Storage Design: Data Volumes, Logs and Backup
HANA separates three storage roles, and each has different requirements:
| Role | Requirement | Recommended build |
|---|---|---|
| Data volumes | High capacity, good read performance | NVMe U.2 array (RAID 5/10 via PERC H965i) или 16 САС + 8 NVMe mix |
| Redo log | Lowest latency, write-dense | Dedicated NVMe, RAID 1, sized ~512GB-1TB |
| Резервное копирование | Емкость + throughput, not latency | External HBA to backup storage; 3-2-1 rule with offsite copy |
On the R960, МИНУТА H965i (tri-mode) covers RAID and NVMe passthrough, while the external HBA355e/HBA465e option extends to backup targets — keep backup I/O off the data path entirely.
Network and High Availability
SAP HANA System Replication (HSR) requires a dedicated low-latency path between primary and secondary — typically a second 10/25/100GbE fabric. A production-grade HA pair is: two R960 scale-up nodes, HSR in synchronous mode, a quorum/ensembling mechanism, and automatic failover tested in advance. The 12 Gen5 slots give each node room for dual fabrics (client + replication) plus HCA without compromise.
Migration Path: Getting to HANA on R960
Moving an existing SAP landscape (any DB) to HANA on a new 4-socket platform follows a proven sequence:
- Size the instance — current data + growth + headroom; confirm it stays under the 16 TB scale-up envelope.
- Order a certified build — memory density first, then NVMe layout, then HA pair with HSR.
- Load and validate — HANA database migration option (DMO) or classic load; run the standard SAP performance checks on the target.
- Cut over — schedule the shortest downtime window; use delta load to minimize it.
- Operate — iDRAC telemetry + HANA Studio monitoring; keep firmware and HANA revision aligned.
The payoff of scale-up: one pair of 4-socket nodes replaces a multi-node scale-out cluster for typical enterprise OLTP — lower license surface, one HA domain, dramatically simpler operations.
Buying Checklist
- Confirm the SAP-certified configuration with SAP’s certified hardware list for the exact Xeon generation and memory.
- Maximize memory per slot (128GB/256GB RDIMM) — 64 slots at 16 TB is the density play.
- Dedicate NVMe for redo logs; keep backups on separate external storage.
- Plan dual fabrics from day one — client network + HSR replication.
- Buy the HA pair together and validate failover before go-live.
FAQ
- 16 TB enough? For the large majority of enterprise OLTP and mixed workloads, yes — beyond that, scale-out or HANA Cloud become the conversation.
- Why 4 sockets over 2? 64 Слоты DIMM (16 ТБ) против 32 (8 ТБ) — memory ceiling is the scale-up differentiator.
- NVMe or SAS for data? NVMe U.2 for performance tiers, SAS/SATA for capacity; the R960 mixes both.
- HA required? Production HANA should run System Replication — it is the difference between RPO≈0 and a restore.
- Can we start scale-out later? Да — but sizing correctly at purchase avoids a costly re-platform.
Planning a HANA refresh or migration? We supply certified builds on the PowerEdge R960 and smaller footprints on the 670 рэндов, with custom configuration and free consultation from our authorized partner team — factory-direct pricing, 3-год гарантии, глобальная доставка. Contact us for a HANA sizing review.