Синькуань Сервер | Поставщик оборудования для корпоративных серверов
The RAID Conversation That Starts Every Storage Architecture Meeting
It is 2026 and storage architects still argue about hardware RAID versus software RAID as if one is objectively correct. The truth is workload-dependent: a PERC H965i hardware RAID controller is the right answer for a Windows Server with local disks, a single-storage-array SQL Server, or an ESXi host using local storage. ZFS, мдадм, and Ceph are the right answer for Linux-native environments, software-defined storage, and hyperconverged clusters. This article explains the architectures, the tradeoffs, and the decision framework that determines which one belongs in your deployment.
The Three RAID Architectures in 2026
| Architecture | How It Works | Example | CPU Overhead | Характеристики производительности |
|---|---|---|---|---|
| Hardware RAID (RAID-on-Chip controller) | Dedicated ASIC/controller card with its own CPU, DRAM cache, and XOR engine. RAID logic runs entirely on the card. OS sees a single virtual disk. | Dell PERC H965i, H755, LSI/Broadcom 9400 ряд | Zero (controller handles all parity, XOR, кеш) | Best for write-heavy with battery/capacitor-backed cache. Predictable latency. Controller cache accelerates writes. |
| Программный RAID (OS-based) | CPU executes RAID logic (XOR, parity, зеркальное отображение) using kernel drivers. OS sees the raw drives and manages redundancy itself. | Linux mdadm, Windows Storage Spaces, ZFS RAID-Z, Btrfs RAID | Moderate (1-2 cores for parity RAID 5/6 at high IOPS; near-zero for RAID 1/10) | RAID 1/10: near line-rate. RAID 5/6: CPU-bound for parity computation but modern CPUs handle it well. |
| HBA Passthrough (no RAID) | Controller passes drives through to OS as individual devices. No RAID logic at all. OS manages drives directly (usually with ZFS or Ceph). | Dell PERC H965i in eHBA mode, LSI 9300 HBA, Бродком 9500 in IT mode | Highest (filesystem manages all redundancy, dedup, compression) | ZFS: excellent (ARC cache + checksums + снимки). Ceph: software-defined replication at scale. |
Hardware RAID: When the Controller Carries the Load
Hardware RAID shines in three specific scenarios:
- Write-heavy workloads with cache: A PERC H965i with 8 GB NV cache (capacitor-backed) absorbs write bursts before destaging to disks. SQL Server transaction logs and VMware swap files benefit from the write cache because the OS issues writes and immediately receives acknowledgement — the controller handles the actual disk write asynchronously.
- Boot volumes and OS drives: An ESXi host booting from a PERC RAID 1 array of two SATA SSDs needs zero OS-level storage management. The controller presents a single bootable disk. No kernel modules, no mdadm configuration, no filesystem-level redundancy to debug.
- Predictable latency: Hardware RAID delivers consistent I/O latency because the controller’s cache and XOR engine handle variance. For workloads with strict latency SLAs (OLTP databases, payment processing), predictable latency is worth the controller cost.
Программный RAID: When the OS Knows Better
Software RAID wins in environments where the filesystem’s features matter more than the controller’s cache:
- ZFS RAID-Z (Linux/FreeBSD/TrueNAS): ZFS provides checksums on every block, silent data corruption detection (a capability hardware RAID lacks), снимки, compression, and deduplication. RAID-Z1/2/3 is software RAID 5/6/7 with integrity verification. The data-integrity story of ZFS is the reason it dominates TrueNAS deployments.
- Linux mdadm: The battle-tested software RAID implementation. RAID 0/1/5/6/10 across any mix of drives. Zero controller dependency — drives move between servers freely. mdadm is the default for Linux distributions for good reason: it is free, гибкий, and reliable.
- Windows Storage Spaces: Microsoft’s software RAID with storage tiering (NVMe + HDD in one pool), storage spaces direct (S2D) for clusters, and ReFS filesystem integration. The Windows-native alternative to hardware RAID controllers.
Tri-Mode Controllers: The Bridge That Made the Choice Less Binary
The Dell PERC H965i and Broadcom 95xx series are tri-mode controllers: they support SAS, САТА, and NVMe drives on the same card, in the same backplane, managed by the same controller. They can operate in RAID mode (аппаратный RAID) or HBA/eHBA mode (passthrough). This single-card flexibility means the decision is no longer «buy a RAID controller or buy an HBA» — it is «buy a tri-mode controller and configure it for the architecture you need today, knowing you can reconfigure it later.»
| Контроллер | Порты | Protocols | RAID Modes | Кэш | eHBA Mode |
|---|---|---|---|---|---|
| МИНУТА H965i | 16 (4x SlimSAS x4) | SAS 12G, SATA 6G, NVMe Gen4/5 | 0, 1, 5, 6, 10, 50, 60 | 8 GB NV (capacitor-backed) | Да |
| ПЕРК H755 | 16 (4x SlimSAS x4) | SAS 12G, SATA 6G, NVMe Gen3/4 | 0, 1, 5, 6, 10, 50, 60 | 8 GB NV | Да |
| LSI 9500-16e | 16 (внешний) | SAS 12G, SATA 6G, NVMe Gen4 | 0, 1, 10 (entry), 5/6 (with key) | 4 ГБ (optional module) | Да (IT mode) |
| Broadcom 9500-8i | 8 (внутренний) | SAS 12G, SATA 6G, NVMe Gen4 | 0, 1, 10, 5, 6, 50, 60 | 4 ГБ | Да (IT mode) |
The Decision Framework: Which RAID Architecture for Your Workload
| Рабочая нагрузка / Environment | Recommended Architecture | Reason |
|---|---|---|
| Windows Server, local disks, SQL Server/Oracle on DAS | Hardware RAID (МИНУТА H965i, RAID 10 for data, RAID 1 for OS) | Write cache accelerates DB logs. OS sees one virtual disk. No Linux filesystem complexity. |
| VMware ESXi, local storage, small clusters | Hardware RAID (МИНУТА H965i, RAID 5 или 10) | ESXi manages local storage as datastores; hardware RAID provides the redundancy layer beneath. |
| Линукс, ZFS, TrueNAS, Proxmox | HBA passthrough (eHBA mode), ZFS RAID-Z2 | ZFS needs direct drive access for checksums and smart error handling. Hardware RAID hides drives from ZFS. |
| Ceph / vSAN / S2D (hyperconverged) | HBA passthrough, software-defined replication | Replication and erasure coding happen at the cluster layer. Hardware RAID would fight the software layer. |
| Большие данные / аналитика (Hadoop, Spark) on local disks | JBOD with HBA passthrough or RAID 0 (data is replicated at app layer) | HDFS replicates data 3x at the application layer. RAID redundancy is unnecessary overhead. |
| СМИ / archive / backup targets (sequential write) | Hardware RAID 6 или RAID 60 (МИНУТА H965i) for redundancy, or ZFS RAID-Z2 for integrity | Sequential write benefits from controller cache; rebuild time matters at 20TB+ drive sizes. |
Three Mistakes to Avoid in 2026
- Putting hardware RAID in front of ZFS or Ceph. ZFS cannot see the individual drives it needs for checksum verification when they are hidden behind a RAID controller. Ceph OSDs behind hardware RAID produce performance anomalies because the controller cache and the cluster’s replication logic conflict. Use eHBA/IT mode for both.
- Using hardware RAID 5 on 20TB+ drives. A single 20TB drive failure triggers a 28-hour rebuild window. The probability of a second failure during rebuild is not negligible. RAID 6 or RAID-Z2 is the minimum for 12TB+ drives; RAID 6 is recommended for all HDD arrays above 8TB per drive.
- Relying on hardware RAID for data integrity. Hardware RAID controllers detect failed drives but cannot detect silent data corruption — a flipped bit that passes checksum. ZFS, Btrfs, and ReFS with checksums catch what hardware RAID misses. If your data matters, run checksumming filesystems on top of (or instead of) аппаратный RAID.
Source Your Storage Infrastructure Through Xincuan
We supply Dell PERC H965i/H755 controllers, Broadcom HBAs, Dell PowerEdge and xFusion FusionServer servers, and the full range of enterprise SSDs and HDDs with factory-direct pricing, 3-год гарантии, и доставка по всему миру. Our storage architects can help you choose the right RAID architecture for your workload.
Request a storage architecture consultation
Explore: МИНУТА H965i | SSD | жесткие диски