NETWORK CONCEPTS
Learn how link capacity, traffic, latency, routing and loss shape delivery - and how application limits and protection affect the result.
01 / CAPACITY
One number rarely describes the whole experience. Start by asking whether it describes capacity, achieved delivery, accumulated data or delay.
Bandwidth means link capacity in this guide. Throughput is the rate actually achieved at a stated measurement point. Transfer is the volume moved over a period, while latency describes delay. Each answers a different question.
A small interactive request can depend heavily on round-trip delay and application response time. A large download needs sustained throughput. A continuous stream also needs delivery to remain consistent over time.
A port rating applies to an interface. A test result applies to particular endpoints and conditions. A transfer allowance applies to a service and accounting period. Compare like quantities before choosing infrastructure.
02 / TRANSFER
A port describes link capacity. Throughput describes the rate delivered at a particular layer; traffic describes data accumulated over time.
An uplink connects a service toward its upstream network. Labels such as 100 Mbps, 1 Gbps, 10 Gbps, 40 Gbps and 100 Gbps describe link capacities. These are general networking examples, not a list of tiers available on every service.
Application delivery can be constrained by protocol overhead, the remote endpoint, routing, congestion, CPU, storage, software or service policy. A 10 Gbps interface does not guarantee a 10 Gbps download. There is no universal efficiency percentage.
Metered services count transfer against an allowance or charging model. Unmetered describes how transfer is accounted for; the connection and other resources still have limits. Check the applicable period, direction and service terms.
| Quantity | Unit | What it describes |
|---|---|---|
| Link rate / throughput | Mbps or Gbps | Bits per second at the stated measurement boundary |
| Byte rate | MB/s or GB/s | Bytes per second; distinguish raw conversion from useful delivery |
| Transfer volume | GB or TB | Accumulated data over a stated period |
Illustrative calculation
bytes/s = bits/s ÷ 8
1,000 Mbps ÷ 8 = 125 MB/s
Lowercase b means bits; uppercase B means bytes. One byte contains eight bits. Mbps is not MB/s, and Gbps is not GB/s. This decimal conversion excludes protocol overhead and is not a throughput promise. MiB and GiB use binary prefixes based on powers of 1,024.
Illustrative calculation
volume in bytes = average bit/s × elapsed seconds ÷ 8
20 Mbps × 2,592,000 seconds ÷ 8 = 6.48 TB
This assumes a constant whole-period average over an illustrative 30 days, using decimal TB. A peak rate is not a monthly average. Actual billing periods and traffic-counting directions depend on the service.
03 / QUALITY
Capacity describes how much can move per second. Delay, delay variation and loss describe other aspects of delivery.
Latency is delay between specified events. One-way delay and round-trip time are different measurements. Distance, route selection, transit and peering paths, queues and processing affect network delay; connection setup and application work also affect the response a user sees.
Jitter commonly means variation in packet delay. Voice, live media and other real-time applications need to accommodate uneven arrival times. Interpret the measurement method and interval rather than assume a universal acceptable threshold.
Loss means packets did not arrive within the measurement criteria. TCP can retransmit missing data, adding work and delay. Real-time media may instead show missing content or use application-specific recovery; late data may no longer be useful.
04 / PATH
Traffic crosses networks whose operators make their own routing decisions. The selected path need not be the shortest line on a map.
A route determines where traffic is forwarded toward its destination. A hop is a forwarding step visible to a particular observation. An ASN identifies an autonomous routing system; it is not a guarantee of geography or performance.
Transit provides purchased reachability through another network. Peering exchanges traffic for agreed reachable destinations between networks, directly or through an exchange. Peering can be paid or settlement-free; it does not automatically provide reachability to the whole Internet.
BGP exchanges inter-network reachability and path information. Operators apply policies when selecting routes. Costs, preferences, reachability and conditions can affect the result, and the return path can differ from the outward path.
05 / PATH
A second link is useful only in the context of the failures it can withstand. Two connections may still share a cable route, device or power source.
One link leaves that connection and its dependencies as potential single points of failure. Recovery needs a working alternative and an operational plan.
Two links through one conduit, device or upstream dependency may fail together. Link count alone does not establish independent failure domains.
Separate physical paths and independent dependencies can withstand more failure scenarios. Validate device and power dependencies, surviving capacity, detection and routing convergence.
06 / DIAGNOSTICS
Choose an observation that matches the symptom. A network probe, an application request and a throughput test answer different questions.
Observes: echo replies, round-trip time and missing replies.
Useful for: repeatable endpoint reachability symptoms.
Does not prove: application availability or bandwidth.
Common misinterpretation: treating ICMP response handling as the health of all traffic.
Observes: probe replies along parts of a route.
Useful for: comparing visible paths.
Does not prove: that a slow intermediate reply delays forwarded traffic.
Common misinterpretation: treating a silent router as end-to-end loss; filtering, response limits and return paths affect observations.
Observes: repeated route and response measurements.
Useful for: investigating persistent destination symptoms over time.
Does not prove: the cause of a problem.
Common misinterpretation: blaming an intermediate hop without downstream or application corroboration.
iPerf measures throughput between cooperating endpoints under controlled conditions. Record protocol, direction, duration and concurrency; the endpoints themselves can limit the result. It does not certify application performance. A browser speed test likewise describes its own test path, not a server’s universal capacity.
07 / PROTECTION
Mitigation capacity describes a protection capability. The service uplink describes normal delivery capacity; the two do not have to be the same size.
Mitigation examines traffic and can reject unwanted load before allowed traffic reaches the service. The uplink and application still constrain that allowed traffic. A large protection figure is not an equally large server port or a guarantee that every attack is filtered.
L3 concerns network packets and addressing; L4 concerns transport behavior and connections; L7 concerns application requests. Bits per second, packets per second and requests per second describe different workloads and cannot be treated as interchangeable capacity numbers.
Network protection addresses traffic on the delivery path. A reverse proxy fronts supported application traffic and needs appropriate origin-access controls. The architecture label alone does not establish coverage of every layer, protocol or attack, and neither approach is universally better.
08 / WORKLOAD PLANNING
Estimate simultaneous delivery demand from concurrent sessions and the bitrate actually served. Total daily viewers do not describe the busiest moment.
A brief download burst and continuous media delivery impose different demands. Account for variable bitrate peaks, overhead and retransmissions as well as the average payload. Check whether storage I/O and processing or transcoding can supply the output.
An origin can deliver directly or feed an optional CDN or edge layer. Distribution can change origin demand, but cache misses, cacheability and live delivery behavior still matter. Model the intended arrangement rather than assume every viewer connects to the origin.
Illustrative calculation
payload bit/s ≈ concurrent sessions × mean delivered bitrate
250 concurrent sessions × 6 Mbps = 1.5 Gbps
This assumes the sessions are served by the origin at the same mean bitrate. It excludes variable bitrate peaks, protocol overhead and retransmissions. Distribution and cache behavior may change origin load; measure those conditions before choosing headroom.
09 / WORKLOAD PLANNING
Upgrade the resource that limits useful work. A larger link cannot make a busy application, slow disk or constrained remote endpoint deliver faster by itself.
Correlate application response time and achieved throughput with CPU, storage, database, TLS and transcoding work. Include the client and remote path. A single transfer and many parallel transfers may expose different constraints.
Some spare capacity can absorb demand variation, growth or a maintenance event. Set the reserve from measured peaks, operating scenarios and uncertainty. There is no universal percentage that fits every workload.
Record the endpoints, direction, duration, protocol and concurrency of each test. Repeat representative work and compare with a normal baseline. A momentary peak is not evidence of sustainable application delivery.
Illustrative calculation
seconds = data bytes × 8 ÷ sustained bit/s
500 GB × 8 ÷ 500 Mbps = 8,000 seconds = 2 h 13 min 20 s
This uses decimal units and assumes a sustained useful rate. If 500 Mbps is only the nominal port ceiling, the result is an ideal lower bound. Storage, protocol work and live changes can increase the actual time.
10 / WORKLOAD PLANNING
Use a concise capacity brief for planning and a repeatable evidence bundle for troubleshooting. Keep requirements separate from conclusions.
TURN UNDERSTANDING INTO REQUIREMENTS
Bring your workload, traffic pattern and operating requirements together before comparing infrastructure.
