IT hardware review: racks, servers and storage

Jamal's strongest area is the equipment in the racks, which is also the area where overconfidence costs the most marks. Facility exams test racks, servers and storage at a broad, practical level: what the parts are, how they are installed, how they fail and how they are replaced safely. IT credentials such as CompTIA Server+ go deeper. This review covers both levels, following DCT-130 and DCT-310, and gives the calculations for rack space, weight, RAID capacity and failure rates that exams like to ask.

Racks and rack units

A standard equipment rack has a mounting width of 19 inches between the rails, set by EIA-310, and its height is measured in rack units (U) of 1.75 inches (1.75 in). A typical cabinet is 42U, 45U or 48U tall; Stonebridge uses 48U cabinets in Halls B to D. Open Rack designs from the Open Compute Project use a wider 21-inch bay and a slightly taller 1.89 in OpenU (DCT-130 lesson 2).

Rack load ratings come in three kinds: static (standing still on its leveling feet), rolling (moving on its casters) and shipping (moving loaded in its crate). The rolling rating is usually the lowest, which is why heavy racks are often moved empty or lightly loaded and filled in place. A fully loaded AI rack can be close to its static rating and well above many rolling ratings.

Server anatomy

A server has the same parts as a desktop computer, built for reliability and remote management: one or two CPU sockets, memory in channels, storage bays, PCIe slots for network cards and accelerators, redundant hot-swappable power supplies and fans, and a baseboard management controller (BMC) for remote access (DCT-310 lessons 1 to 4).

• Memory is installed in channels, and the maker's population rules decide which slots are filled first and in what combinations. Unbalanced memory still works but runs slower; mixing types that the board does not allow may stop it from booting.
• Redundant power supplies are usually 1+1. Each supply should go to a different feed, A and B, so that losing one feed does not drop the server. A server with both supplies on the same feed has no power redundancy at all.
• Fans are usually N+1 inside the server; a failed fan raises the speed of the others and logs an event.
• GPU servers and trays for AI carry several accelerators joined by high-speed links; in Hall D's liquid-cooled racks, cold plates on the GPUs and CPUs carry most of the heat to the coolant (DCT-310 lesson 3).
• The BMC runs even when the server is off, as long as it has power. Through it, staff read sensors and logs, power the server on and off, and see the console. Because it can control the server, the BMC must be on an isolated management network with unique credentials (DCT-310 lesson 7).

Storage

Hard disk drives (HDDs) store data on spinning platters; solid-state drives (SSDs) on flash memory. NVMe SSDs connect directly over PCIe and are much faster than drives on SATA or SAS. Drives are grouped for capacity and protection in RAID arrays or, in larger systems, spread across many servers with software protection such as erasure coding (DCT-310 lessons 5 and 6).

Exam traps on RAID: RAID 0 has no redundancy; RAID is not a backup, because deleting a file deletes it on every drive at once; and during a rebuild after one failure, a RAID 5 array has no protection left, which is why large drives are usually put in RAID 6. Bayview Credit Union's RAID 6 array in Hall A was mid-rebuild in DCT-310, and the rule there was no further drive pulls until the rebuild finished.

Failure rates

Drive makers quote mean time between failures (MTBF), and fleet operators measure annualized failure rate (AFR). For a component with a constant failure rate, AFR ≈ 8,760 / MTBF in hours. A drive with an MTBF of 2,000,000 hours has an AFR of about 0.44 percent, but real fleets often see more: Backblaze reported an overall AFR of about 1.36 percent across its drive fleet for 2025. Expected failures per year in a fleet are simply AFR times the number of drives, which is how spares are planned (DCT-310 lesson 11).

# IT hardware review: RAID usable capacity and expected drive failures
def raid_usable(level, n, size_tb):
    if level == 0:  return n * size_tb
    if level == 1:  return size_tb
    if level == 5:  return (n - 1) * size_tb
    if level == 6:  return (n - 2) * size_tb
    if level == 10: return n * size_tb / 2
for level, n in ((0, 8), (5, 8), (6, 8), (10, 8)):
    print(f"RAID {level:<2} with {n} x 16 TB: {raid_usable(level, n, 16):5.1f} TB usable")

mtbf_h = 2_000_000
afr = 8760 / mtbf_h
print(f"MTBF {mtbf_h:,} h -> AFR {afr * 100:.2f} %")
fleet = 4800
for rate in (afr, 0.0136):
    print(f"{fleet} drives at AFR {rate * 100:.2f} % -> about {fleet * rate:.0f} failures a year")

RAID 0  with 8 x 16 TB: 128.0 TB usable
RAID 5  with 8 x 16 TB: 112.0 TB usable
RAID 6  with 8 x 16 TB:  96.0 TB usable
RAID 10 with 8 x 16 TB:  64.0 TB usable
MTBF 2,000,000 h -> AFR 0.44 %
4800 drives at AFR 0.44 % -> about 21 failures a year
4800 drives at AFR 1.36 % -> about 65 failures a year

Replacing hardware safely

The replacement procedure is the same for every part, and exams test its order (DCT-310 lessons 9 and 10):

• Confirm the ticket: the right customer, cabinet, U position, serial number and part. Most hardware mistakes are the wrong server, not the wrong technique.
• Check the server's state through the BMC or with the customer: is it safe to pull this part now? For drives, is the array healthy apart from this drive?
• Use the locator light to find the exact server, and check the label against the ticket.
• Wear an ESD strap, handle parts by their edges, and keep new parts in their bags until fitted.
• Replace the part, confirm it is seen and healthy, and record the old and new serial numbers.
• Return the failed part under RMA with the right paperwork, or destroy it on site if it holds data and the customer's rules require it.

Exam traps on hardware

• 1U = 1.75 inches; 42U is about 2 m of mounting space, not the cabinet's outside height.
• Both power supplies of a server must go to different feeds for power redundancy.
• RAID is not a backup; RAID 0 is not redundant; RAID 5 is exposed during rebuild.
• AFR from MTBF assumes a constant failure rate; new and very old drives fail more often (the bathtub curve, lesson 9).
• The BMC is powerful: isolate its network and never leave default passwords.
• Hot-swappable means designed to be replaced while running, not that any part may be pulled at any time.

Rack dimensions and loading rules

A rack unit (U) is 1.75 inches high, and a standard 42U rack has 73.5 inches of mounting space. Equipment is 19 inches wide between the mounting rails, while the cabinet itself is usually about 24 inches wide, or about 31 inches when extra room is needed for cabling. Depth varies from about 39 to 47 inches, and deep GPU servers may need the deepest cabinets.

Cabinets have two weight ratings. The static rating is what the cabinet can hold standing still on its leveling feet; the dynamic or rolling rating is what it can hold while being moved on its casters, and it is always lower. Exams like the trap of moving a fully loaded cabinet across the floor: unless the rolling rating covers the load, the equipment comes out first. Heavy equipment goes at the bottom to keep the center of gravity low, stabilizers or anti-tip feet go on before any server is pulled out on its rails, and only one server is extended at a time.

Remote management

Every modern server has a baseboard management controller (BMC), a small computer on the motherboard with its own network port that runs even when the server is off. Through the BMC, an administrator can power the server on and off, see its hardware health, view its console and light its locator. Older tools use IPMI; newer ones use the DMTF Redfish interface. Technicians at Stonebridge rarely log in to customer BMCs, but they read BMC alerts in tickets and use the locator lights that customers turn on for them (DCT-310 lesson 3).

Storage interfaces and media

Storage can be direct-attached (DAS, inside or cabled to one server), network-attached (NAS, sharing files over the network) or a storage area network (SAN, giving servers block storage over a dedicated network such as Fiber Channel or over Ethernet). An exam may ask which one lets several servers share block storage: the SAN.

RAID rebuild risk

RAID protects against drive failure, but a rebuild is a period of risk. While a RAID 5 array rebuilds, it has no redundancy left, and every remaining drive must be read in full. With large drives, a rebuild can take a day or more, and a second failure or an unreadable sector in that time can lose data. That is why large arrays use RAID 6 or erasure coding, why hot spares start rebuilds at once, and why a technician never pulls a second drive from a degraded array, even if its light looks odd.

Firmware and compatibility

Replacement parts must match what the server and the customer support. A drive with the wrong firmware may not join the array; a power supply of a different wattage may cause the server to raise a mismatch alarm or limit its power; a DIMM in the wrong slot may stop the server from booting or reduce memory speed. Read the part number and the ticket, not just the size and shape, and follow the vendor's population rules for memory. Firmware updates are changes and follow change management.

Drive sanitization and chain of custody

Drives that leave a customer's control carry data. NIST SP 800-88 describes three levels of sanitization: clear (overwriting with standard commands), purge (methods such as cryptographic erase or block erase that defeat laboratory recovery) and destroy (shredding, disintegrating or pulverizing). Customers choose the level, and many forbid drives from leaving the site at all, so failed drives are kept in a locked bin and shredded under witness, with a certificate for each serial number. A technician's job is to keep the chain of custody unbroken: who removed the drive, when, where it was stored and who destroyed it. Returning a customer's failed drive to a vendor under an RMA when the contract says keep and destroy is a serious breach.

High-density GPU servers

The servers in AI halls bring their own rules. An eight-GPU server can weigh well over 220 lb and draw on the order of 10 kW, so four of them fill a rack's power budget in a few units of space. They need a lift and two people, more power cords than older servers, high-speed cables with tight bend limits, and in Hall D, liquid connections that are checked before the server is powered. Exams set in modern facilities increasingly ask about these limits.