Why Measure Web Performance?

Peak-hour congestion, peering bottlenecks, slow name servers, sluggish servers, and TCP behaviour on long or lossy paths all slow a connection down, and link utilisation counters do not show it.

Such measurements provide figures that can be compared across regions, access technologies and providers and used in SLAs and regulatory reports.

What the User Waits For

A single web or file transfer passes through several phases. Each one can be measured on its own, and each points to a different part of the chain.

Phases of a Transfer
PhaseWhat it shows
DNS resolutionWhether name servers are reachable and fast
Connection setupRound-trip time to the server and server responsiveness
Time to first byteHow quickly the service starts delivering
Download and upload throughputThe capacity available to an application
Transfer successReliability: completed, failed and aborted transfers, with the reason for each failure or abort

From Network Tests to Service Tests

Web performance can be tested on several complementary layers.

Network benchmarking
IETF RFC 2544 defines throughput, latency, frame loss and back-to-back tests, originally for benchmarking network devices in the lab. ITU-T Y.1564 is its multiservice successor for service activation: it verifies several traffic classes at once against their committed rates, delay, delay variation and loss. x-fabric's Coast Guard benchmarks IP networks to both.
TCP throughput
Most of the traffic runs over TCP, and TCP throughput depends on round-trip time, window size and loss, not just on link capacity. IETF RFC 6349 describes a repeatable method: determine path MTU, baseline round-trip time and bottleneck bandwidth first, then measure TCP throughput and compare it with what the path should achieve. Metrics such as TCP efficiency and buffer delay show why a line underperforms. Coast Guard measures TCP throughput this way.
Application transfers
HTTP and FTP downloads and uploads against a test server measure what a customer application gets. Broadband Forum TR-143 standardises such throughput tests (HTTP and FTP transfers, UDP echo) for customer premises equipment. ETSI TS 102 250-2 defines QoS parameters for popular services, including FTP and HTTP, with exact trigger points for setup time, session time, mean data rate and failure ratios, so that results from different tools and networks are comparable. The mosprobe-ng measures at this layer.
User perception
ITU-T G.1030 provides a framework for estimating how IP network performance translates into the perceived performance of data applications such as web browsing.

One Connection or Many?

A single TCP connection is sensitive to latency and loss. Several parallel connections fill even fast lines.

Utilisation: 80% with 1 connection

One connection. What a single application gets.

Utilisation: 92% with 4 connections

Parallel connections. The total capacity of the access line.

Running uploads and downloads at the same time reveals effects that sequential tests miss, such as growing latency when buffers fill under load.

What a Good Measurement Needs

The customer's vantage point
Measure from real access lines, not from the data centre.
The right targets
A reference server close to the network edge isolates the access network; real internet targets bring peering, content delivery networks and DNS into the measurement.
Both directions and both IP versions
Upload and download, IPv4 and IPv6 measured separately.
Long enough transfers
For short transfers, TCP slow start dominates the result rather than the capacity of the line.
A test head that is not the bottleneck
Gigabit lines need test equipment that reliably sustains gigabit rates.
Around the clock
Performance at 3 a.m. says little about performance at peak hours.
Context
DNS time, connection setup and time to first byte, together with round-trip time and loss to reference targets, explain why a transfer was slow.