LAN performance indicators may be grouped into fixed, variable, and performance measurement metrics [TERP96].

  Fixed metrics

•  Transmission capacity — The transmission capacity is normally expressed in terms of bits/second. Although the bite rate is fixed, the total capacity can be divided into multiple smaller capacities to support different types of signals. One of the common myths regarding LAN transmission capacity is that Ethernet is saturated at an offered load (the actual data carried on the channel, excluding overhead and retransmitted bits) of 37%. Many studies have shown that Ethernet can support a 10 Mbps data rate under a distance of one kilometer with the CSMA/CD protocol.

•  Signal propagation delay — Signals are limited by the speed of light, and the longer they propagate, the longer the delay. Signal propagation time is the time required to transmit a signal to its destination and generally is 5 microseconds per kilometer. Therefore, cabling distance is a factor that affects signal propagation delay. In the case of satellite communication, signal propagation delay plays an influential role, as the distance between an earth station and the satellite is about 22,500 miles. Within LANs, the internodal signal propagation delay is negligible. However, the signaling technique used (e.g., base-band or broadband) can produce different levels of delays.

•  Topology — A LAN can be a star, tree, ring, bus, or combination of star and ring. The type of LAN topology will affect performance. For example, a bus LAN (e.g., Ethernet) and a token ring LAN (e.g., IBM’s token ring) have a different built-in slot time — the time of acquiring network access. The topology also limits the number of workstations or hosts that can be attached to it. Ethernet limits the number of nodes per cable segment to 100, and the total number of nodes in a multiple-segment Ethernet is limited to 1024. A single IBM token ring supports 260 nodes. The higher the number, the greater the performance impact since all network traffic is generated from these nodes.

•  Frame/packet size — Most LANs are designed to support only a specific, fixed size of frame or packet. If the message is larger than the frame size, it must be broken into smaller sizes occupying multiple frames. The greater the number of frames per message, the longer the delay a message can experience. Like every other LAN, Ethernet, for example, has a minimum packet size requirement: it must not be shorter than the slot time (51.2 microseconds) in order to be able to detect a collision. This limit is equivalent to a minimum length of 64 bytes, including headers and other control bytes. Similarly, Ethernet has a maximum of 1518 bytes as the upper boundary, in order to minimize access time.

  Variable metrics

•  Access protocol — The type of access protocol used by a LAN is probably the most influential metric that affects performance. IBM’s token ring uses a proprietary token access control scheme, in which a circulating token is passed sequentially from node to node to grant transmissions. A node must release a token after each transmission and is not allowed to transmit continuously on single ring architecture. Ethernet, on the other hand, employs the I-persistent CSMA/CD access control in which a node that waits for a free channel can transmit as soon as the channel is free with a probability of 1 (i.e., 100% chance to transmit).

•  User traffic profile — A computer system and network is lifeless without users. Many factors constitute a user’s traffic profile: message/data arrival rate (how many key entries a user makes per minute), message size distribution (how many small, medium, and large messages are generated by a user), type of messages (to a single user, multiple users, or all receivers), and the number of simultaneous users (all active, 50% active, or 10% active.)

•  Buffer size — A buffer is a piece of reserved memory used to receive, store, process, and forward messages. If the number of buffers is too small, data may suffer delays or be discarded. Some LANs have a fixed number of buffers, and some use a dynamic expansion scheme based on the volume of the messages and the rate of processing. In particular, LAN internetworking devices are likely sources of buffer problems.

•  Data collision and retransmission — Data collision is inevitable, especially in a bus LAN, unless the transmission is controlled in an orderly manner. Two factors need to be considered: how long it takes nodes to detect a data collision and how long it takes to actually transmit the collided messages. Various detection schemes are used by different topologies. For example, Ethernet employs a “jam” time, which is the time to transmit 32 to 48 more bits after a collision is detected by a transmitting station so that other stations can reliably detect the collision. The more influential factor is the time to actually transmit the data after collision. Many LANs use a binary exponential backoff scheme to avoid a situation in which the same two colliding nodes collide again at the next interval. Both collision detection and retransmission contribute delays to the overall processing delay. Generally, waiting time is dependent on network load and may become unacceptably long in some extreme cases.

The performance of a LAN cannot be quantified with a single dimension. It is very hard to interpret measured metrics without knowing what applications (users) are involved. The following measurement metrics are generally obtainable:

•  Resource usage — Processor, memory, transmission medium, and in some cases, peripheral devices all contribute to the processing of a user request (e.g., open a file, send a message, or compile a program). How much of their respective capacities are used and how much reserved capacities are left need to be evaluated in conjunction with processing delay information (in some cases, user’s service level goals).

•  Processing delays — A user’s request is likely to suffer delays at each processing point. Both host and network can cause processing delays. Host delays can be divided into system processing and application processing delays. Network delays can be viewed as a combination of delays due to hardware and software. However, at the end user level, a total processing delay (or response time) is the only meaningful performance metric.

•  Throughput — Transmission capacity can be measured in terms of throughput — the number of messages or bytes transmitted per unit of time. In LAN measurement, throughput is an indication of the fraction of the nominal network capacity that is actually used for carrying data. In general, packet headers are considered useful in estimating throughput, if no other measurement facilities are available, since the header contains the number of bytes in a frame. A metric related to throughput is channel capacity. Each transmission medium has a given maximum capacity (e.g., bits/second), which is a function of message volume and message size.

•  Availability — From an end user’s point of view, service availability is determined by its availability and consistency. A network can be in operation, but if a user suffers long delays, as far as the user is concerned, the network is virtually unavailable since it is seen as unreliable. Therefore, reliability measurement is a permanent measurement metric. However, most LAN measurement tools are only able to measure availability (up and down time), since timing measurement may add several orders of magnitude of complexity to measurement tools

•  Fairness of measured data — Since network traffic tends to be sporadic, the measured period and the internal data-recording rate are quite important. An hourly averaged measured data rate may not be able to reveal any performance bottlenecks; a 1-second recording rate can generate an enormous amount of data that requires both processor time and storage. As a general practice, a peak-to-average ratio is used in which data in short intervals with known high activity are collected. The ratio between the high activity periods and the average periods can be established for studying network capacity requirements.

•  Communigram — In order to quantify the traffic between communication partners, the volume is quite important. The measured and reported intervals are very important. An hourly averaged rate may not be able to reveal any performance bottleneck; a 1-second recording rate can generate an enormous amount of data that requires both processor time and storage. As a general practice, a peak-to-average ratio is used in which data in short intervals with known high activity are collected. The ratio between the high activity periods and the average periods can be used for sizing resources supporting the communication between partners.