Standard server racks are 19 inches wide (EIA-310 standard) between mounting rails.
Common heights:
42U: the most common full-height data center rack, about 6 ft / 2 m tall
45U–48U: taller racks for high-density data centers
24U–27U: half-height racks for small offices and server rooms
6U–12U: wall-mount racks for small networking gear
Common depths range from 600 mm to 1200 mm. Servers usually need deep racks (1000 mm or more) to fit the chassis and rear cabling.
Rack styles:
2-post (relay rack): open frame, for lightweight network gear such as switches and patch panels
4-post open frame: supports servers with rails, but has no doors or side panels
Enclosed cabinet: lockable doors and side panels for physical security and controlled airflow
1 rack unit (U) = 1.75 inches (44.45 mm) of vertical space.
Each U has 3 mounting holes on each rail.
Typical uses by size:
1U (“pizza box”): dense web servers and network appliances. Limited expansion slots and drive bays, and small fans that run loud.
2U: the most common general-purpose server. More drive bays, PCIe slots and better cooling.
3U–4U: storage servers, GPU servers and systems with many expansion cards.
Rail hole types:
Square holes: use cage nuts or tool-less rails. The most common type.
Round holes: unthreaded, for specific rail kits.
Threaded holes: screws go in directly. Common on 2-post racks.
Exam tip: to work out whether equipment fits, add up the U heights of every device. A 42U rack holding a 4U UPS, two 1U switches, a 1U KVM drawer and a 1U patch panel has 35U left for servers.
Place the heaviest equipment at the bottom: UPS units, battery packs and large storage arrays.
Place lighter equipment higher: switches, patch panels and other network gear, usually at the top near the cable trays.
Put servers in the middle.
Place the KVM console at a comfortable working height, roughly waist to chest level.
Plan for growth by leaving empty U spaces, and cover unused spaces with blanking panels.
Label every device and record its rack position (for example “Rack 3, U15–U16”) in documentation or a Data Center Infrastructure Management (DCIM) tool.
Mount all equipment facing the same direction so airflow is consistent, usually front intake and rear exhaust.
Hot aisle / cold aisle: racks are arranged in rows so fronts face fronts (cold aisle) and backs face backs (hot aisle). Cool air is pulled in at the front of each server and hot air is exhausted out the back.
Containment: physical barriers enclose the hot aisle or the cold aisle so hot and cold air do not mix.
Blanking panels: fill empty U spaces so hot exhaust air does not recirculate to the front intakes.
Raised floors and perforated tiles: deliver cold air from CRAC/CRAH units up into the cold aisle.
CRAC/CRAH: Computer Room Air Conditioner and Computer Room Air Handler units.
ASHRAE recommended intake temperature: about 18–27 °C (64–81 °F). Humidity should also be controlled. Air that is too dry increases Electrostatic Discharge (ESD) risk, and air that is too humid causes condensation.
Keep cable bundles from blocking airflow, especially at the rear of servers.
Use temperature and humidity sensors to monitor the rack, and watch for hot spots.
Bend at the knees, not at the waist, and keep your back straight.
Hold the load close to your body and avoid twisting while carrying it.
Use team lifts for heavy equipment. Many organizations require two people for anything over about 50 lb (23 kg).
Use a server lift or mechanical lift for heavy servers, especially when mounting in the upper U positions.
Remove power supplies and drives to reduce weight before lifting, if needed.
Wear appropriate PPE and keep pathways clear.
Load from the bottom up, with the heaviest items lowest, to keep the center of gravity low.
Extend the anti-tip stabilizer feet or bolt the rack to the floor before installing or sliding out equipment.
Only extend one device at a time on its rails. Pulling several out at once can tip the rack.
Distribute weight evenly from side to side and front to back.
Lock the caster wheels and level the rack with the leveling feet.
A fully loaded 42U rack can weigh 1,000–3,000 lb (450–1,350 kg).
Check the floor’s load rating, especially on raised floors and in upper floors of buildings. The relevant ratings are:
Static load: the weight the floor holds while the rack is stationary
Rolling load: the weight it can bear while a rack is moved across it
Concentrated (point) load: the weight on a single caster or foot
Remove perforated floor tiles carefully, and do not roll heavy racks over unsupported tiles.
Consult facilities or structural engineers before adding high-density racks.
A PDU distributes power from a circuit or UPS to the devices in a rack. It is essentially an enterprise-grade power strip.
Types:
Basic: distributes power only, with no monitoring.
Metered: displays the total current draw on the unit itself.
Monitored: reports power usage remotely over the network.
Switched (managed): allows remote on/off control and power cycling of individual outlets.
Automatic transfer switch (ATS) PDU: switches between two input sources if one fails.
Mounting:
Vertical (0U): mounts in the rear side channels of the rack and uses no rack units. This is the most common in data centers.
Horizontal: uses 1U or 2U of rack space.
Use two PDUs (A and B feeds) per rack for redundancy, with each fed from a separate circuit or UPS.
Do not overload a PDU. Follow the 80% rule and keep continuous load at or below 80% of the circuit rating.
A KVM switch lets one keyboard, monitor and mouse control multiple servers.
A rack-mounted KVM console drawer (usually 1U) has a fold-out LCD screen and keyboard.
Place it at a comfortable working height, typically in the middle of the rack, around U20–U25 in a 42U rack.
IP KVM (KVM over IP): provides remote console access over the network, including BIOS-level access. It complements out-of-band management such as iDRAC, iLO and IPMI.
Cascade KVM switches when there are many servers, and label each KVM port so it maps to the correct server.
Rails mount the server in the rack and are usually supplied by the server vendor.
Types:
Sliding rails: let the server slide out for maintenance without removing it from the rack. The most common type.
Static (fixed) rails or shelves: the server cannot slide out. Cheaper, and used for lighter equipment.
Tool-less rails: snap directly into square-hole racks.
Cable management arm (CMA): attaches to the rear of sliding rails so cables fold and extend with the server without being pulled loose.
Installation tips:
Confirm the rail kit supports the rack’s hole type and depth.
Make sure both rails sit in the same U position and are level.
Mount the inner rails on the server, then slide the server onto the outer rails until it locks.
Test that the server slides freely and that the safety latches engage.
Dual (redundant) power supplies: servers have two or more hot-swappable PSUs.
Connect each Power Supply Unit (PSU) to a different Power Distribution Unit (PDU) (A feed and B feed), with each fed from a different circuit or UPS. If one feed fails, the server stays up on the other.
Redundancy models:
N+1: one spare unit beyond what is required.
2N: a fully duplicated power path.
Connecting both PSUs to the same PDU defeats the redundancy, and this mistake is a common exam scenario.
A UPS provides battery backup during an outage and conditions incoming power, protecting against surges, sags and spikes.
Types:
Standby (offline): switches to battery when the power fails. Lowest cost, with a brief transfer time.
Line-interactive: regulates voltage (AVR) without using the battery. Common in small server rooms.
Online (double-conversion): always runs from the inverter, so there is zero transfer time and the best protection. Standard in data centers.
Sizing:
Ratings are given in VA (volt-amps) and watts. Calculate the total load and allow headroom.
Runtime depends on the load and battery capacity. A UPS is meant to bridge the gap until a generator starts or until systems shut down gracefully.
Connect the UPS to servers by USB, serial or a network management card so they can shut down gracefully during long outages.
Mount the UPS at the bottom of the rack because the batteries are heavy.
Test it regularly and replace batteries every 3–5 years.
Feed the A and B PDUs from different electrical circuits or breakers, ideally from different electrical panels.
This prevents a single tripped breaker from taking down a server.
Use two different utility feeds or power providers where possible. This protects against a utility-level outage.
Data centers combine this with generators and an automatic transfer switch (ATS). The typical chain is utility → ATS → generator backup → UPS → PDU → server PSU.
C13 / C14: the most common server connector, rated up to about 10 A (15 A in North America).
C14 is the male inlet on the server PSU.
C13 is the female plug on the cord that goes into the PSU.
C19 / C20: higher current, about 16–20 A. Used for high-power servers, blade chassis and large PDUs.
C5 / C6 (“cloverleaf” or “Mickey Mouse”): used on laptops and small devices.
C7 / C8 (“figure 8”): small two-wire devices.
Typical PDU-to-server cord: C13 at the server end, C14 at the PDU end.
NEMA 5-15: the standard US 120 V / 15 A wall outlet.
NEMA 5-20: 120 V / 20 A, with one T-shaped slot.
NEMA L5-30: twist-lock, 120 V / 30 A.
NEMA L6-30: twist-lock, 208/250 V / 30 A. Common for PDUs and UPS units.
NEMA L14-30: twist-lock, 120/240 V / 30 A.
Twist-lock connectors (the “L” designation) lock into place so they cannot be pulled out accidentally.
IEC 60309 (pin and sleeve): industrial, heavy-duty connectors for high-capacity PDUs. The color indicates voltage, for example blue for 230 V and red for 400 V.
Server PSUs typically support 100–240 V auto-ranging. 208/230 V is more efficient than 120 V in data centers.
Color-code the A and B feeds, for example red for A and blue for B.
Use cords of the correct length to avoid excess slack.
Use cord retention clips or locking cords so cables cannot be pulled out accidentally.
Separate power cables from data cables to reduce EMI. Route power down one side of the rack and data down the other.
Label both ends of each power cable with the PDU and outlet number.
Use hook-and-loop (Velcro) straps rather than tight zip ties.
Do not block fan exhausts or PSU airflow.
Use multiple NICs or ports per server.
NIC teaming (bonding) combines ports for failover and/or load balancing. LACP (IEEE 802.3ad / 802.1AX) is the standard link aggregation protocol.
Connect each NIC to a different switch so the server survives a switch failure.
Use redundant uplinks, redundant core switches and multiple paths.
Storage multipathing (MPIO) provides redundant paths to SAN storage.
Uses RJ-45 connectors and the T568A or T568B wiring standard.
The maximum channel length for most categories is 100 meters (328 ft).
Categories:
Cat 5e: 1 Gbps up to 100 m.
Cat 6: 1 Gbps up to 100 m, and 10 Gbps up to about 55 m.
Cat 6a: 10 Gbps up to 100 m. The common data center choice for 10GBASE-T.
Cat 7: 10 Gbps, shielded. Not an official TIA standard.
Cat 8: 25/40 Gbps up to 30 m. Used for short top-of-rack links.
UTP (unshielded) vs STP / F/UTP (shielded): shielded cable resists EMI and must be properly grounded.
Plenum-rated cable (CMP) is required in air-handling spaces because it produces less toxic smoke when it burns.
Fiber transmits light instead of electrical signals.
Advantages: immune to EMI, longer distances, higher bandwidth, and more secure because it is hard to tap.
Disadvantages: more fragile, sensitive to bending, and more expensive.
SC (Subscriber or Square Connector):
Square, push-pull snap-in design with a 2.5 mm ferrule.
Larger than LC. Common in older installations and telecom.
LC (Lucent Connector):
Small form factor with a 1.25 mm ferrule and a latch mechanism.
Often used as a duplex pair. The standard connector for SFP and SFP+ transceivers and most common in modern data centers.
Small core of about 9 microns.
Uses a laser light source with a single light path.
Carries signals over long distances, from kilometers up to 40 km or more.
Jacket is typically yellow.
More expensive optics. Used between buildings, campuses and WAN links.
Larger core of 50 or 62.5 microns.
Uses LED or VCSEL light sources, and multiple light paths travel through the core.
Shorter distances, up to a few hundred meters.
Grades:
OM1/OM2: orange jacket.
OM3/OM4: aqua jacket. Carry 10 Gbps up to about 300 m (OM3) and 400 m (OM4).
OM5: lime green jacket.
Cheaper optics. Most common inside data centers.
Do not exceed the minimum bend radius.
Keep connectors clean and capped because dust causes signal loss.
Never look directly into a fiber or optic, since the laser can cause eye damage.
Single-mode and multimode optics and cables must not be mixed.
1000BASE-T: copper, Cat 5e or better, up to 100 m.
1000BASE-SX: multimode fiber, short reach (about 550 m).
1000BASE-LX: single-mode fiber, long reach (about 5–10 km).
10GBASE-T: copper, Cat 6a up to 100 m (Cat 6 up to about 55 m).
10GBASE-SR: multimode fiber, short reach (300–400 m).
10GBASE-LR: single-mode fiber, long reach (10 km).
10GBASE-CR (DAC, twinax): direct attach copper, short runs of about 1–7 m. Used for in-rack and top-of-rack connections.
A hot-swappable modular transceiver that plugs into a switch or NIC port.
Lets one port support different media, such as copper, multimode or single-mode fiber.
Rated for 1 Gbps. Usually uses LC fiber connectors, though copper RJ-45 SFPs also exist.
Same physical size as SFP, but supports 10 Gbps.
An SFP+ port can usually accept a 1 Gbps SFP, but an SFP port cannot run SFP+ at 10 Gbps.
Related: SFP28 supports 25 Gbps.
Has four channels in one module and is physically larger than SFP.
Speeds:
QSFP: 4 × 1 Gbps = 4 Gbps
QSFP+: 4 × 10 Gbps = 40 Gbps
QSFP28: 4 × 25 Gbps = 100 Gbps
Breakout cables can split one QSFP+ port into four 10 Gbps SFP+ connections.
Usually uses MPO/MTP multi-fiber connectors, or LC for some types.
DAC (direct attach copper): a twinax cable with transceivers built on each end. Cheap, low-power and short-range.
AOC (active optical cable): fiber with transceivers built on each end, for longer runs.
Some vendors lock their switches to vendor-coded optics, so check compatibility before buying.
Label both ends of every cable and document port mappings.
Use patch panels and horizontal and vertical cable managers.
Color-code cables by function, for example management, production, storage or uplinks.
Respect bend radius, especially for fiber.
Leave a small service loop for maintenance without excess slack.
Use Velcro, not tight zip ties, because overtightening can damage the cable and degrade performance.
Keep data cables away from power cables.
Do not block airflow, and use cable management arms on sliding rails.
A standalone, upright chassis that looks like a large desktop PC.
Pros: low cost, quiet, easy to deploy with no rack needed, and lots of internal expansion space.
Cons: uses a lot of floor space, scales poorly, and makes cable management harder.
Best for: small businesses, branch offices, and sites with only one or two servers.
Some towers can be converted to rack mount with a conversion kit.
Mounts in a standard 19-inch rack using rails, and is sized in U (1U, 2U, 4U).
Pros: high density, centralized cabling, power and cooling, and easy to scale.
Cons: requires a rack and infrastructure, and is louder.
Best for: data centers and server rooms. This is the most common enterprise chassis.
A chassis that holds multiple thin blade servers, typically about 6–10U tall and holding 8–16 blades.
Blades share the chassis’s power supplies, cooling fans, network and storage switches or interconnect modules, and management modules.
Pros:
Highest density.
Fewer cables.
Centralized management.
Hot-swappable blades.
Cons:
High initial cost.
Vendor lock-in, because blades work only in the same vendor’s chassis.
High power and cooling demand per rack.
The chassis can become a single point of failure if not configured redundantly.
Best for: large virtualization clusters and high-density compute environments.
A vendor-published list of hardware tested and certified to work with a particular OS, hypervisor or server platform. Examples include VMware’s compatibility guide, Microsoft’s Windows Server catalog and Red Hat certified hardware.
Always check the HCL before buying or installing components such as CPUs, RAM, RAID controllers, NICs and drives.
Using non-HCL hardware can cause driver problems or instability, and can void vendor support or warranty.
Also check the server manufacturer’s list of supported parts, including approved DIMMs, drives and option cards.
Main vendors: Intel Xeon and AMD EPYC. ARM server processors are also used, for example in cloud environments.
Key specifications:
Cores and threads (hyper-threading / SMT)
Clock speed
Cache
TDP (power and heat)
Number of memory channels
PCIe lanes
Socket type must match the motherboard, for example Intel LGA or AMD SP-series sockets.
Multi-socket servers: all CPUs must normally be the same model and stepping. Populate CPU sockets in the order the vendor specifies.
Each CPU controls its own memory slots (NUMA). Adding a second CPU often enables additional DIMM slots and PCIe slots.
Virtualization extensions (Intel VT-x / AMD-V) must be enabled in the BIOS/UEFI for hypervisors.
Installation:
Use ESD protection.
Align pin 1 and do not bend the socket pins.
Apply thermal paste correctly and seat the heatsink evenly.
Update the BIOS/UEFI firmware first if the new CPU requires it.
Server uses:
AI/ML training and inference
High-performance computing (HPC)
VDI (virtual desktops with GPU acceleration)
Video rendering and transcoding
Considerations:
Power: high-end GPUs draw 300–700 W and need supplemental power cables.
Physical size: often full-height, full-length, double-width cards that may need a 2U–4U chassis.
Cooling: data center GPUs are often passively cooled and depend on the server’s chassis fans, so the server must support them.
PCIe slot: typically a full x16 slot.
Check the HCL and server vendor support, as well as drivers and licensing (for example for vGPU).
GPU passthrough or vGPU lets virtual machines use the GPU.
Servers use ECC (error-correcting code) memory, which detects and corrects single-bit errors.
DIMM types:
UDIMM (unbuffered): lower capacity. Used in entry-level servers.
RDIMM (registered/buffered): has a register chip that reduces electrical load, allowing more DIMMs and higher capacity. The most common server memory.
LRDIMM (load-reduced): allows even higher capacity per server.
Do not mix UDIMM, RDIMM and LRDIMM types in the same server.
DDR generations (DDR3, DDR4, DDR5) are physically keyed differently and are not interchangeable.
Population rules:
Follow the vendor’s slot population order to use all memory channels.
Use matched DIMMs (same size, speed and rank) for best performance.
Populate evenly across CPUs in multi-socket systems.
Mixing speeds causes all memory to run at the speed of the slowest module.
Advanced features:
Memory mirroring: duplicates data across channels for fault tolerance and halves usable capacity.
Memory sparing: reserves a spare rank in case a module fails.
Handle memory with ESD protection, and seat the module fully until the clips lock.
PCIe (PCI Express): the current standard expansion bus. It uses point-to-point serial lanes.
Slot widths: x1, x4, x8, x16. More lanes means more bandwidth.
Approximate bandwidth per lane, per direction:
PCIe 3.0: about 1 GB/s
PCIe 4.0: about 2 GB/s
PCIe 5.0: about 4 GB/s
Each generation doubles the speed of the previous one, and PCIe is backward compatible.
A smaller card works in a larger slot, for example an x4 card in an x16 slot.
A slot may be physically x16 but electrically wired as x8, so check the motherboard documentation.
Legacy buses:
PCI: parallel and shared.
PCI-X: a faster server version of PCI. Older exam material may still refer to it.
Processor interconnects link CPUs in multi-socket systems: Intel QPI/UPI and AMD Infinity Fabric.
SATA:
Up to 6 Gbps.
Low cost, used for bulk and lower-tier storage.
SAS (Serial Attached SCSI):
12 Gbps (SAS-3), with 24 Gbps SAS-4 also available.
Enterprise-grade, dual-ported for redundancy, and more reliable.
SAS controllers and backplanes can run SATA drives, but SATA controllers cannot run SAS drives.
NVMe:
SSDs that connect directly over PCIe lanes, giving the highest performance and lowest latency.
Form factors: M.2, U.2/U.3 (2.5-inch hot-swappable drives) and add-in cards.
SCSI (parallel): a legacy interface.
USB: peripherals and boot media. Often disabled for security.
Ethernet: built-in LOM/NIC ports and the dedicated out-of-band management port (iDRAC, iLO, IPMI/BMC).
Fibre Channel: SAN connectivity through an HBA, at 8/16/32/64 Gbps.
Serial/console port: legacy management access.
Common server expansion cards:
NIC: additional or faster network ports, such as 10/25/40/100 GbE.
HBA (host bus adapter): connects to SAN storage (Fibre Channel) or external SAS storage.
RAID controller: hardware RAID with cache, ideally with battery- or flash-backed cache.
GPU: compute and graphics acceleration.
NVMe adapter: adds NVMe drives through a PCIe slot.
Riser cards: change slot orientation or add slots in 1U and 2U servers, where cards are installed horizontally.
Physical sizing:
Full-height vs low-profile (half-height): swap the bracket to match the chassis.
Full-length vs half-length: make sure there is physical clearance.
Check the HCL and vendor compatibility.
Verify the slot has enough lanes and generation for the card, for example a 100 GbE NIC needs a PCIe x16 Gen3/Gen4 slot.
Confirm power, cooling and physical space, and whether a riser is required.
Power down, unplug, use ESD protection, and seat the card firmly.
Install drivers and firmware and update them to supported versions.
Verify the card in the BIOS/UEFI and the OS, then test functionality.
Some slots are tied to a specific CPU in multi-socket servers. If that CPU is not installed, the slot will not work.
Tipping hazard when sliding out a server → extend the stabilizer feet, load heavy items at the bottom, and extend only one device at a time.
Server lost power even though it has dual PSUs → both PSUs were probably on the same PDU or circuit. Use A/B feeds on separate circuits.
Hot spots or overheating in the rack → check blanking panels, hot/cold aisle orientation, and cables blocking airflow.
10 Gbps over copper at 90 m → Cat 6a.
Link between buildings 5 km apart → single-mode fiber.
40 Gbps uplink → QSFP+. 100 Gbps → QSFP28.
Short connection between a server and its top-of-rack switch → DAC (twinax).
New RAID card or NIC not recognized or unstable → check the HCL, firmware and drivers, and slot compatibility.
New memory not detected or performance poor → check the DIMM type (RDIMM vs UDIMM), the population order, and the DDR generation.
Maximum compute density with shared power and cooling → blade enclosure.
Small office with one server and no rack → tower.
Remote BIOS-level console access across many servers → IP KVM or out-of-band management (iDRAC, iLO, IPMI).