Selecting the Right Server
A server is not a single product — it is a set of choices. The same budget can buy a machine that flies or one that chokes, depending on whether the CPU, memory, storage, RAID, and network match the workload. This lesson turns “get a server” into a specification you can defend.
Start with the workload
Every server decision flows from what the machine will actually do. Before you look at a part list, collect four inputs:
Users and devices
How many clients depend on this server, and what each one asks of it. A file open and a database query are worlds apart.
Peak concurrent connections
Not the average. A server that is idle at 10 a.m. and saturated at 9 a.m. must be sized for 9 a.m.
Storage needs
Usable capacity today, plus how fast data must be read and written — input/output operations per second, or IOPS.
Growth
The expected yearly increase. A good rule for infrastructure is 3–5 years of headroom, because servers are usually refreshed on that cycle.
The classic mistake is over-buying CPU while under-buying storage or memory. Most real servers stall on I/O or run out of RAM long before the processor is the bottleneck.
CPU: cores versus clock speed
A processor has cores (independent execution units) and a clock speed in gigahertz (GHz). More cores let the CPU run many tasks at once; a higher clock lets each task run faster. A server may hold one or two sockets, each holding one CPU package, and most modern cores support simultaneous multithreading (Intel Hyper-Threading), presenting two logical processors per core.
| Workload | What it stresses | CPU profile |
|---|---|---|
| Web / file server | Many small, mostly independent requests | More cores, moderate clock |
| Database | Few complex queries, heavy on cache and single-thread speed | Fewer, faster cores + large cache |
| Virtualization host | Runs many virtual machines at once | Very high core count + lots of RAM |
| Batch / analytics | Long CPU-bound jobs that parallelise well | Maximum cores |
Also watch core-based software licensing. Database and virtualization platforms are often licensed per physical core, so adding cores can raise the software bill as fast as the hardware one. That is a justification input, not a specification detail, but it changes the answer more often than students expect.
Memory: capacity and ECC
RAM holds the working data and the operating system. On a server, memory is rarely just “how many gigabytes” — it is also what kind.
- Capacity. Fit enough RAM that the working set stays in memory and the server is not constantly paging to disk. Virtualization multiplies this: if each VM needs 8 GB and you plan to run 10, budget at least 80 GB before the hypervisor.
- ECC memory. Error-Correcting Code memory detects and corrects single-bit errors and detects multi-bit errors in transit. A consumer PC can tolerate a rare flipped bit; a server holding a database or a bank ledger cannot, because a silent corruption can cascade into wrong results or a crash. ECC is the default in server platforms.
- Registered (buffered) DIMMs. Server memory is often RDIMM, which places a register between the memory chips and the controller. This improves stability and lets the board address far more memory than unbuffered UDIMM allows.
- Population rules. Memory runs fastest when DIMMs are spread across all channels. A board with eight slots and two channels may want DIMMs in pairs; an odd or unbalanced population can halve memory bandwidth.
Storage: HDD, SSD and NVMe
Storage is where most servers feel slow or fast, so learn the three canonical media and what each is for:
- HDD (hard disk drive) — magnetic platters that spin and a mechanical head that seeks. It is cheap per terabyte and high capacity, but slow (millisecond access), because the head must physically move. Best for cold data, archives, and bulk files where latency does not matter.
- SSD (solid-state drive) — flash memory with no moving parts. It is far faster than HDD (microsecond access, thousands of IOPS) and rugged, at a moderate price per terabyte. Typically connects over SATA or SAS. The mainstream choice for operating systems and general server workloads.
- NVMe (Non-Volatile Memory express) — flash driven directly over PCIe instead of a SATA/SAS controller. It is the fastest option: the highest IOPS and lowest latency, ideal for databases, caching tiers, and any workload that is IOPS-bound. It also costs the most per terabyte.
RAID by workload
RAID (Redundant Array of Independent Disks) combines several physical disks into one logical volume for speed, protection, or both. Learn the four levels by their purpose:
| RAID level | Technique | Gives you | Cost |
|---|---|---|---|
| RAID 0 | Striping across disks | Speed; no redundancy — one disk loss loses all data | 100% capacity |
| RAID 1 | Mirroring (identical copies) | Safety: survives one disk failure | 50% capacity |
| RAID 5 | Striping + distributed parity (min 3 disks) | Safety + capacity: survives one disk failure | One disk’s worth lost |
| RAID 10 | Mirror + stripe (min 4 disks) | Speed + safety: survives a failed disk per mirror | 50% capacity |
Guidance by workload: RAID 0 only for scratch or caches that can be rebuilt; RAID 1 for operating systems and small critical volumes; RAID 5 for general file and application data where capacity plus single-disk protection matters; RAID 10 for databases and virtual machines where both speed and safety are required.
Networking and form factor
How the server joins the network is part of the specification.
- NIC speed. Current server network interface controllers run at 1, 10, 25, or 40 GbE (and beyond). A file server feeding hundreds of clients, or a virtualization host moving VM images, needs 10 GbE or higher; a small branch server may be fine at 1 GbE. Look for dual-port NICs so you can team them.
- NIC teaming. Also called bonding or link aggregation, teaming joins two or more physical NICs into one logical link. It adds bandwidth (traffic spreads across the links) and redundancy (if one link or switch fails, the others carry on).
- Remote management. Out-of-band management lets you control the server even when the OS is down: IPMI is the open standard, iLO is HPE’s implementation, and iDRAC is Dell’s. These give a virtual console and power control over a separate network port — essential for a headless rack server.
Form factor: rack versus tower
The physical shape follows the room it will live in:
- Rack-mounted — measured in rack units (U), where 1U is 44.45 mm tall. A 1U server is thin and dense; a 2U server allows more drives, cards, and better airflow. Racks centralise power, cooling, and cabling and are the norm in data centres and server rooms.
- Tower — a stand-alone chassis that sits on the floor or a shelf. It needs no rack, runs quieter, and suits a small office with one or two servers. It is harder to scale and wastes vertical space.
- Blade — compute modules that slide into a shared chassis. Very high density, but a bigger up-front investment and shared components to consider.
Worked example: a 200-user file server
The situation. A 200-employee office needs a central file server. Users store office documents, spreadsheets, and images. Peak load is around 80 concurrent users at 9 a.m. Usable data is about 20 TB and grows roughly 15% a year. The server will live in a small rack in the server room.
Walk the four subsystems:
- CPU. File serving is I/O-bound, not CPU-bound. A single mid-range processor with around 16 cores is ample and leaves room to add a virtual machine later. No need for a top-tier database chip.
- Memory. Fit 64 GB ECC. The file service benefits from a large read cache, and 64 GB comfortably covers the OS, the caching, and one or two future VMs without paging.
- Storage. Put the OS on a RAID 1 pair of SSDs so the system volume is fast and survives a disk failure. Put the bulk data on a RAID 5 array of high-capacity HDDs sized for roughly 20 TB usable plus the parity disk and growth — for example, four 8 TB drives give about 24 TB usable. Add an SSD read cache (or a small NVMe tier) so frequently opened files feel snappy despite the spinning bulk tier. If budgets were tighter on performance instead of capacity, RAID 10 on SSDs would be the alternative.
- Networking. Two 10 GbE ports, teamed, so throughput scales and a single link failure does not take the server offline. Enable iDRAC or iLO out-of-band management for headless administration.
Form factor: a 2U rack server, which holds the drive bays and offers better cooling than a cramped 1U chassis. The result is a spec that is defensible because every choice maps to a stated requirement — not to a brand or a budget number.
Selection table
Use this as a starting point, then adjust with the four inputs from Part 1:
| Workload | CPU | RAM | Storage | RAID |
|---|---|---|---|---|
| Web / app server | Many cores, moderate clock | 32–64 GB ECC | SSD (OS + app) | 1 (OS), 5 or 10 (data) |
| Database server | Few fast cores, big cache | Large, ECC | NVMe + fast SSD | 10 |
| File server | Moderate cores | 64 GB ECC | SSD OS + HDD bulk + cache | 1 (OS), 5 (data) |
| Virtualization host | Max cores | Very large, ECC | SSD/NVMe shared store | 10 |
| Backup / archive | Low | 16–32 GB ECC | High-capacity HDD | 5 or 6 |
Scenario drills
A start-up runs a public website on one server and expects traffic to triple. Which subsystem should they upgrade first, and why?
A client wants the cheapest possible storage for a backup archive that is written once and rarely read. What do you recommend?
A database server must survive a disk failure and still deliver high IOPS. Which RAID level and media fit best?
Why do servers use ECC memory when desktops often do not?
A rack server will run headless in a data centre. What feature lets you access it when the OS is unreachable?
Review
Check what you have learned with these review questions.
Sources & references
- Cisco Networking Academy. CCNA: Introduction to Networks (ITN), module resources on network devices and infrastructure. netacad.com.
- Cisco. CCNA (200-301) Exam Topics — Network Fundamentals and IP Services. cisco.com.
- Microsoft Learn. Windows Server storage and hardware guidance (storage spaces, RAID alternatives, and memory requirements). learn.microsoft.com.
- Dell Technologies. iDRAC documentation. dell.com.
- HPE. iLO documentation. hpe.com.
Target exams: CCNA 200-301 v1.1 · CyberOps Associate 200-201. Last content verification: 2026-10-11.