RAID (Redundant Array of Independent Disks) combines multiple physical drives into one logical unit. Depending on the level, this improves performance, fault tolerance, or both.
Key terms:
Striping: data is split into blocks and spread across multiple drives. This improves performance.
Mirroring: identical data is written to two or more drives. This provides redundancy.
Parity: calculated data (XOR) that allows lost data to be rebuilt if a drive fails.
Hot spare: an idle drive that automatically replaces a failed drive and starts the rebuild.
Rebuild: the process of reconstructing data onto a replacement drive. Performance drops during a rebuild, and the array is vulnerable until it completes.
Write penalty: the number of back-end disk operations needed for each write the server makes. Higher write penalties mean slower writes.
RAID is not a backup. It protects against drive failure, but not against deletion, corruption, ransomware, or the loss of the whole array.
Minimum drives: 2
Usable capacity: 100% (the sum of all drives)
Fault tolerance: none. If one drive fails, all data is lost.
Performance: fastest read and write, with no write penalty.
Use cases: temporary data, scratch space, video editing cache, or anything easily recreated.
Example: 4 × 2 TB drives = 8 TB usable.
Minimum drives: 2
Usable capacity: 50%
Fault tolerance: survives the failure of one drive in a two-drive mirror.
Performance: good reads, since either drive can serve them. Writes are about the same speed as a single drive (write penalty of 2).
Use cases: OS and boot volumes, and small critical data sets.
Rebuilds are simple and fast because the data is just copied.
Example: 2 × 1 TB drives = 1 TB usable.
Minimum drives: 3
Usable capacity: (N − 1) × drive size, where one drive’s worth of space holds parity.
Fault tolerance: survives one drive failure.
Performance: good reads. Slower writes, with a write penalty of 4 (read data, read parity, write data, write parity).
Use cases: general file servers and read-heavy workloads.
Weaknesses:
Rebuilds on large drives take a long time and stress the remaining drives.
An unrecoverable read error (URE) or a second failure during a rebuild loses the array.
For this reason RAID 5 is not recommended for large high-capacity HDDs.
Example: 4 × 2 TB drives = 6 TB usable.
Minimum drives: 4
Usable capacity: (N − 2) × drive size.
Fault tolerance: survives two simultaneous drive failures.
Performance: good reads, but the slowest writes of the common levels (write penalty of 6).
Use cases: large-capacity arrays, archives, backup targets, and arrays built from large drives where long rebuilds are a risk.
Example: 6 × 4 TB drives = 16 TB usable.
Minimum drives: 4, and always an even number.
Usable capacity: 50%
Fault tolerance: survives one drive failure per mirrored pair. It can survive several failures as long as both drives in the same pair do not fail.
Performance: excellent read and write performance (write penalty of 2), with fast rebuilds.
Use cases: databases, virtualization hosts, email servers, and other write-intensive, high-I/O workloads.
The trade-off is cost: half the raw capacity is used for redundancy.
Example: 8 × 2 TB drives = 8 TB usable.
Drives are presented either individually or spanned/concatenated into one large volume.
No striping, no mirroring and no parity, so there is no fault tolerance and no performance gain.
Usable capacity: 100%, and drives of different sizes can be combined.
If a drive fails, the data on that drive is lost. On a spanned volume, the whole volume may be lost.
Use cases: maximizing capacity cheaply, and software-defined storage that handles redundancy itself, such as ZFS, Ceph or Windows Storage Spaces.
“JBOD” also describes an external drive enclosure with no RAID controller, connected to a server through a Serial Attached SCSI (SAS) Host Bus Adapter (HBA).
RAID 0: minimum 2 drives, 100% capacity, 0 failures tolerated, fastest.
RAID 1: minimum 2 drives, 50% capacity, 1 failure tolerated.
RAID 5: minimum 3 drives, N−1 capacity, 1 failure tolerated, write penalty 4.
RAID 6: minimum 4 drives, N−2 capacity, 2 failures tolerated, write penalty 6.
RAID 10: minimum 4 drives, 50% capacity, 1 failure per mirror pair tolerated, write penalty 2.
JBOD: minimum 1 drive, 100% capacity, 0 failures tolerated.
A dedicated RAID controller card or onboard chip with its own processor handles RAID.
Advantages:
Offloads parity calculations from the CPU.
Usually has cache memory protected by a battery (BBU) or flash module (FBWC) so cached writes survive a power loss.
OS independent: the OS sees a single logical disk.
Can boot from any RAID level.
Supports hot spares and hot-swap drives, and is managed through the controller BIOS or vendor tools (for example Dell PERC or HPE Smart Array).
Disadvantages:
Higher cost.
The controller can be a single point of failure.
The array may only be readable by the same model or family of controller.
Controller firmware must be kept up to date.
The operating system manages RAID. Examples: Windows Disk Management and Storage Spaces, Linux mdadm, and ZFS.
Advantages:
Low or no cost, with no special hardware.
Arrays can move between systems running the same OS.
Flexible.
Disadvantages:
Uses host CPU and memory.
OS dependent.
Booting from some levels can be limited.
Usually has no battery-protected write cache.
RAID provided by the motherboard chipset, which relies on OS drivers to do the work.
A low-cost middle ground found in entry-level servers.
Capacity planning means estimating storage needs now and in the future so the server does not run out of space or performance.
What to consider:
Current usage: how much data exists today.
Growth rate: the historical trend (for example 20% per year) projected across the server’s lifecycle, typically 3–5 years.
RAID overhead: usable capacity is lower than raw capacity. For example, RAID 1/10 loses 50%, RAID 5 loses one drive, and RAID 6 loses two.
File system and formatting overhead, plus the space reserved by the OS.
Snapshots, replicas and backup copies, which all consume extra space.
Data retention requirements: legal and compliance rules for how long data must be kept.
Data reduction: deduplication and compression can lower the required capacity. Results vary with the type of data.
Performance (IOPS and throughput): plan for speed as well as space. Some workloads need more drives or SSDs for IOPS even when capacity is sufficient.
Headroom: plan to stay below roughly 70–80% utilization. Performance degrades and the risk of running out rises as volumes fill.
Thin vs. thick provisioning:
Thick: all space is allocated up front.
Thin: space is allocated as data is written. This is efficient, but overcommitting means a pool can run out unexpectedly, so it must be monitored.
Decimal vs. binary units: drives are sold in decimal TB (10¹² bytes), while operating systems often report binary TiB (2⁴⁰ bytes). A “4 TB” drive shows as about 3.64 TiB.
Monitoring and alerting: set threshold alerts for capacity, and review trends regularly.
Expansion options: free drive bays, external JBOD shelves, SAN/NAS growth, or cloud tiering.
Requirement: 10 TB today, growing 25% per year for 3 years, with fault tolerance for two drive failures.
Year 3 need: 10 × 1.25³ ≈ 19.5 TB. Adding about 25% headroom gives roughly 24–25 TB usable.
RAID 6 with 8 TB drives gives (N − 2) × 8 TB. Using 5 data drives + 2 parity = 7 drives gives 40 TB usable, or 5 drives gives 24 TB usable.
Choose the configuration that meets the requirement with headroom, and add a hot spare.
Uses NAND flash memory and has no moving parts.
Advantages over HDDs:
Much faster random I/O, with high IOPS and very low latency.
Lower power use and less heat.
Silent.
More resistant to shock.
Disadvantages:
Higher cost per GB.
Limited write endurance.
NAND cells can only be written and erased a limited number of times (P/E cycles, or program/erase cycles) before they wear out.
Cell type affects endurance, cost and capacity:
SLC (1 bit per cell): highest endurance and most expensive.
MLC (2 bits per cell): high endurance.
TLC (3 bits per cell): moderate endurance. The most common type today.
QLC (4 bits per cell): lowest endurance, highest capacity and cheapest.
Endurance ratings:
TBW (terabytes written): the total data that can be written over the drive’s life.
DWPD (drive writes per day): how many times the full capacity of the drive can be written each day over the warranty period, typically 5 years.
Features that extend SSD life:
Wear leveling spreads writes evenly across all cells.
Overprovisioning reserves hidden spare capacity for wear leveling and to replace bad blocks.
TRIM / UNMAP tells the SSD which blocks are no longer in use.
Garbage collection reclaims blocks in the background.
Write amplification: the SSD often writes more data internally than the host requested, which speeds wear.
Monitor wear through S.M.A.R.T. attributes (for example “percentage used” or “wearout indicator”) and vendor tools. Replace drives proactively.
Rated for about ≤1 DWPD.
Lower cost and higher capacity, usually TLC or QLC flash.
Use cases: read-heavy workloads such as web servers, content delivery, boot drives, data warehousing, and read caches.
Rated for about 10+ DWPD.
More overprovisioning and higher-endurance NAND.
Use cases: write-heavy workloads such as transactional databases, logging, and write caching or tiering.
More expensive.
Rated for about 3 DWPD.
A balance between read-intensive and write-intensive drives, used for virtualization and general workloads.
Spinning magnetic platters with read/write heads on an actuator arm.
Advantages:
Low cost per GB.
Very high capacities, at 20 TB and above.
Disadvantages:
Mechanical parts that are prone to failure and sensitive to shock and vibration.
Slower random I/O, because of seek time and rotational latency.
Higher RPM means lower rotational latency and more IOPS, but also more heat, more power, higher cost and lower capacity.
About 175–210 IOPS per drive.
Typically 2.5-inch SAS drives with lower capacity.
Used for high-performance enterprise workloads such as databases. Largely being replaced by SSDs.
About 125–150 IOPS per drive.
Typically 2.5-inch SAS drives.
A balance of performance and capacity. Common for enterprise application and VM storage.
About 75–100 IOPS per drive.
Typically 3.5-inch SATA or nearline SAS (NL-SAS) drives.
Highest capacity and lowest cost per GB.
Used for bulk storage, file shares, backups and archives.
5,400 RPM drives also exist, but are mostly found in desktops, laptops and low-power external drives.
Hybrid drive (SSHD): a single HDD with a small built-in flash cache. Frequently accessed data is cached in flash for speed.
Hybrid storage arrays and tiering: arrays combine SSDs and HDDs. Hot (frequently used) data is automatically moved to the SSD tier, and cold data stays on the HDD tier. SSDs can also serve as a cache in front of the HDDs.
The goal is close to SSD performance for active data with HDD cost and capacity for
Enterprise drive interface that uses the SCSI command set over a serial connection.
Speeds: 3, 6, 12 Gbps, with SAS-4 at 22.5 Gbps (marketed as “24G”).
Features:
Full duplex operation.
Dual-port drives, which allow redundant paths to two controllers for high availability.
SAS expanders support many drives, used in JBODs and drive shelves.
Longer cable lengths than SATA, up to about 10 m.
Higher reliability ratings (MTBF).
Nearline SAS (NL-SAS): 7,200 RPM high-capacity drives built on SATA-class mechanics with a SAS interface.
Compatibility: SAS controllers and backplanes accept SATA drives, but SATA controllers cannot run SAS drives. The SAS connector is keyed differently to prevent this.
Lower-cost drive interface, common in desktops and entry-level servers.
Speeds:
SATA I: 1.5 Gbps
SATA II: 3 Gbps
SATA III: 6 Gbps
Half duplex, single port, with a maximum internal cable length of about 1 m.
Supports the AHCI standard, which enables hot-plug and Native Command Queuing (NCQ).
Used for high-capacity, low-cost bulk storage and boot drives.
In a storage context, this means drives and storage controllers attached directly to the PCI Express (PCIe) bus.
NVMe (Non-Volatile Memory Express) SSDs connect over PCIe lanes and avoid the SATA/SAS controller bottleneck. They have very high throughput and IOPS and very low latency.
Form factors:
PCIe add-in cards (AIC)
M.2 (often used for boot drives)
U.2/U.3 (2.5-inch hot-swappable NVMe drives)
EDSFF (E1/E3)
PCIe slots also host RAID controllers, SAS HBAs and Fibre Channel HBAs.
Bandwidth depends on the PCIe generation and number of lanes. For example, a PCIe 4.0 x4 NVMe drive can reach about 8 GB/s.
The external version of SATA, for connecting external drives and enclosures.
Speeds up to 3 or 6 Gbps, about the same as internal SATA.
Shielded cable up to 2 m, with a connector that differs from internal SATA.
Standard eSATA does not supply power, so the external drive needs its own power source. eSATAp (Power over eSATA) combines eSATA and USB power in one port.
Largely replaced by USB 3.x and Thunderbolt.
Used for external drives, backup media, OS installation media and boot devices.
Speeds:
USB 2.0: 480 Mbps
USB 3.0 / 3.1 Gen 1 / 3.2 Gen 1: 5 Gbps
USB 3.1 Gen 2 / 3.2 Gen 2: 10 Gbps
USB 3.2 Gen 2×2: 20 Gbps
USB4: 40 Gbps or more
Hot-pluggable and supplies bus power.
Some servers have an internal USB port for booting a hypervisor or holding a license dongle.
Security: USB ports are often disabled in BIOS/UEFI or by policy to prevent data theft and malware.
Flash memory cards. In servers, these are mainly used to boot embedded hypervisors such as VMware ESXi.
Many servers have dual mirrored SD card modules for redundancy, for example Dell IDSDM and HPE dual SD.
Formats:
SD: up to 2 GB
SDHC: up to 32 GB
SDXC: up to 2 TB
Limitations:
Low write endurance and slow speed.
Not suitable for logs, swap or heavy writes.
VMware has deprecated SD and USB as standalone boot media for ESXi 7 and later because of wear-related failures. Vendors now recommend M.2 boot drives (for example the Dell BOSS card) instead.
Shared storage lets multiple servers access centralized storage. It is required for clustering, VM live migration, and centralized backup and management.
The key distinction is file-level (NAS) versus block-level (SAN).
A dedicated storage appliance or server that shares files over a standard Ethernet network.
File-level access: the NAS manages the file system, and clients access files and folders through shares.
Uses the existing LAN and IP addressing. Easy to set up and manage, and relatively low cost.
Use cases: file shares, home directories, media storage, backup targets, and VM datastores over NFS.
Disadvantages:
Shares bandwidth with LAN traffic unless it is on a separate network or VLAN.
Higher overhead than block storage for some workloads, such as databases.
A file-sharing protocol native to UNIX and Linux, also supported by VMware ESXi datastores.
Uses TCP/UDP port 2049.
Versions:
NFSv3: stateless, with host-based (IP) access control.
NFSv4: stateful, TCP only, and supports Kerberos authentication and better security.
Clients mount exports, for example mount server:/export /mnt/share.
Permissions are set in the export configuration (for example /etc/exports) and through UNIX file permissions.
A dialect of SMB (Server Message Block), the native Windows file-sharing protocol.
Uses TCP port 445. Older NetBIOS-based versions used ports 137–139.
Access uses UNC paths such as \\server\share, with mapped drives and NTFS/share permissions integrated with Active Directory.
CIFS is essentially SMB 1.0, which is outdated and insecure (exploited by WannaCry/EternalBlue). Disable SMBv1 and use SMB 3.x, which supports encryption, multichannel and better performance.
Linux can share files to Windows clients using Samba.
A dedicated high-speed network that gives servers block-level access to shared storage.
Storage is divided into LUNs (Logical Unit Numbers). A server sees each LUN as a local disk, then formats it and manages its file system.
Advantages:
High performance and low latency.
Supports clustering, VM live migration and boot from SAN.
Centralized storage management, snapshots and replication.
Disadvantages:
Expensive and complex, and needs specialized skills.
Key terms:
Initiator: the client, usually the server or its HBA.
Target: the storage array or port.
LUN masking: controls which hosts can see which LUNs, configured on the storage array.
Zoning: controls which devices can communicate across the fabric, configured on the Fibre Channel switch.
Multipathing (MPIO): redundant paths between server and storage, for failover and load balancing.
Sends SCSI commands over TCP/IP Ethernet, which makes it a lower-cost SAN.
Uses TCP port 3260.
Components:
Initiator: either a software initiator built into the OS (Windows iSCSI Initiator, Linux open-iscsi, ESXi) or a hardware iSCSI HBA that offloads processing.
Target: the storage array, which presents LUNs.
Names use IQN (iSCSI Qualified Name) format, for example iqn.2026-11.com.example:storage1.
Security: CHAP (Challenge Handshake Authentication Protocol) authentication, plus network isolation.
Best practices:
Use a dedicated network or VLAN for iSCSI traffic.
Use jumbo frames (MTU 9000) consistently end-to-end.
Use MPIO with multiple NICs and switches.
Use 10 GbE or faster.
Advantages: uses existing Ethernet skills and equipment, and is cheaper than Fibre Channel.
A dedicated, high-performance, lossless SAN protocol and network.
Speeds: 8, 16, 32, 64 Gbps (and 128 Gbps).
Hardware:
FC HBAs in servers.
FC switches, forming the “fabric”.
Usually fiber optic cabling with LC connectors and SFP transceivers.
Addressing uses WWN (World Wide Name): a WWNN identifies a node and a WWPN identifies a port. These are similar to MAC addresses.
Topologies:
Point-to-point: direct connection between a server and storage.
Arbitrated loop (FC-AL): legacy shared loop.
Switched fabric: the most common topology, using FC switches.
Security and access control uses zoning on switches and LUN masking on storage.
Best practice is two separate fabrics (Fabric A and Fabric B) with dual-port HBAs for redundancy.
Advantages: highest performance and reliability, and isolated from LAN traffic.
Disadvantages: expensive, requires specialized hardware and skills, and needs separate infrastructure.
Encapsulates Fibre Channel frames inside Ethernet frames, so LAN and SAN traffic can share one network (converged networking).
Requirements:
10 GbE or faster with lossless Ethernet, provided by Data Center Bridging (DCB) features such as Priority Flow Control.
CNAs (Converged Network Adapters): a single card that acts as both NIC and FC HBA.
FCoE-capable switches.
Not routable over IP, because it runs directly on Ethernet at Layer 2 with no TCP/IP. This is a key difference from iSCSI.
Advantages: fewer adapters, cables and switches, with FC management retained.
Disadvantages: needs specialized DCB-capable equipment, and has seen limited adoption compared with iSCSI and native FC.
NAS (NFS / SMB-CIFS): file-level, Ethernet/IP. Easy and inexpensive.
iSCSI SAN: block-level, Ethernet/IP (TCP 3260). Routable and inexpensive.
Fibre Channel SAN: block-level, dedicated FC network. Fastest and most reliable, and most expensive.
FCoE SAN: block-level, lossless Ethernet at Layer 2. Converged, but not IP-routable.
Database or VM host needs best write performance plus redundancy → RAID 10.
Large-capacity array that must survive two drive failures → RAID 6.
Most usable capacity with single-drive fault tolerance using 3+ drives → RAID 5.
Mirrored boot drives for the OS → RAID 1.
Maximum speed with no redundancy needed → RAID 0.
Mixed-size drives combined for maximum capacity, no redundancy → JBOD.
Need to offload RAID from the CPU and protect cached writes → hardware RAID controller with battery- or flash-backed cache.
SSDs failing early on a write-heavy database → use write-intensive (high-DWPD) SSDs, and monitor wear through S.M.A.R.T.
Cheapest bulk storage for backups or archives → 7,200 RPM SATA or NL-SAS HDDs.
SATA drive in a SAS backplane → works. SAS drive on a SATA controller → does not work.
Linux clients need shared file access → NFS. Windows clients → SMB/CIFS, with SMBv1 disabled.
Servers need storage that appears as a local disk → SAN (block-level): iSCSI, FC or FCoE.
Low-cost SAN using the existing Ethernet infrastructure → iSCSI on a dedicated VLAN with jumbo frames and MPIO.
Host can’t see its LUN on a Fibre Channel SAN → check zoning on the switch, LUN masking on the array, and the WWPNs.
Converge LAN and SAN onto one set of adapters and switches → FCoE with CNAs and DCB.
Hypervisor boot media keeps failing on SD cards → move to mirrored M.2 boot drives.