Network throughput: Difference between revisions

Revision as of 09:13, 26 June 2024

Revision as of 10:05, 26 June 2024

Line 10:

The '''system throughput''' or '''aggregate throughput''' is the sum of the data rates that are delivered to all terminals in a network.<ref>[[Guowang Miao]], Jens Zander, K-W Sung, and Ben Slimane, Fundamentals of Mobile Data Networks, Cambridge University Press, {{ISBN|1107143217}}, 2016.</ref> Throughput is essentially synonymous to [[digital bandwidth consumption]]; it can be determined numerically by applying the [[queueing theory]], where the load in packets per time unit is denoted as the arrival rate ({{mvar|λ}}), and the drop in packets per unit time is denoted as the departure rate ({{mvar|μ}}).

The throughput of a communication system may be affected by various factors, including the limitations of the underlying analog physical medium, available processing power of the system components, [[end-user]] behavior, etc. When taking various protocolsprotocol [[Overhead (computing)|overheads]] into account, the useful rate of the data transfer can be significantly lower than the maximum achievable throughput; the useful part is usually referred to as [[goodput]].

==Maximum throughput==

{{See also|Peak information rate}}

Users of telecommunications devices, systems designers, and researchers ininto communication theory are often interested in knowing the expected performance of a system. From a user perspective, this is often phrased as either "Whichwhich device will get my data there most effectively for my needs?", or "Whichwhich device will deliver the most data per unit cost?". Systems designers often select the most effective architecture or design constraints for a system, which drivesdrive its final performance. In most cases, the benchmark of what a system is capable of, or its "maximum performance" is what the user or designer is interested in. The term maximum throughput is frequently used when discussing end-user maximum throughput tests.  

Maximum throughput is essentially synonymous withto [[digital bandwidth capacity]].

Four different values are relevant in the context of "maximum throughput", used in comparing the 'upper limit' conceptual performance of multiple systems. They are 'maximum theoretical throughput', 'maximum achievable throughput', 'peak measured throughput', and 'maximum sustained throughput'. These values represent different quantities, and care must be taken that the same definitions are used when comparing different 'maximum throughput' values. Each bit must carry the same amount of information if throughput values are to be compared. [[Data compression]] can significantly alter throughput calculations, including generating values exceeding 100% in some cases. If the communication is mediated by several links in series with different bit rates, the maximum throughput of the overall link is lower than or equal to the lowest bit rate. The lowest value link in the series is referred to as the [[bottleneck (traffic)|bottleneck]].

===Maximum theoretical throughput===

This number is closely related to the [[channel capacity]] of the system,<ref>Blahut, 2004, p.4</ref> and is the maximum possible quantity of data that can be transmitted under ideal circumstances. In some cases, this number is reported as equal to the channel capacity, though this can be deceptive, as only non-packetized systems (asynchronous) technologies can achieve this without data compression. Maximum theoretical throughput is more accurately reported taking into account format and specification [[protocol overhead|overhead]] with best- case assumptions. This number, like the closely related term 'maximum achievable throughput' below, is primarily used as a rough calculated value, such as for determining bounds on possible performance early in a system design phase.

===Asymptotic throughput===

The '''asymptotic throughput''' (less formal ''asymptotic bandwidth'') for a packet-mode [[communication network]] is the value of the [[maximum throughput]] function, when the incoming network load approaches [[infinity]], either due to a [[Message passing|message size]],<ref>'' Modeling Message Passing Overhead'' by C.Y Chou et al. in Advances in Grid and Pervasive Computing: First International Conference, GPC 2006 edited by Yeh-Ching Chung and José E. Moreira {{ISBN|3540338098}} pages 299-307</ref> or the number of data sources. As other [[bit rate]]s and [[data bandwidth]]s, the asymptotic throughput is measured in [[bits per second]] (bit/s) or (rarely) [[byte]]s per second (B/s), where 1 B/s is 8 bit/s. [[Decimal prefix]]es are used, meaning that 1&nbsp;Mbit/s is 1000000 bit/s.

Asymptotic throughput is usually estimated by sending or [[network simulation|simulating]] a very large message (sequence of data packets) through the network, using a [[greedy source]] and no [[flow control (data)|flow control]] mechanism (i.e., [[User Datagram Protocol|UDP]] rather than [[Transmission Control Protocol|TCP]]), and measuring the network path throughput in the destination node. Traffic load between other sources may reduce this maximum network path throughput. Alternatively, a large number of sources and sinks may be modeled, with or without flow control, and the aggregate maximum network throughput measured (the sum of traffic reaching its destinations). In a network simulation model with infinite packet queues, the asymptotic throughput occurs when the [[Network latency|latency]] (the packet queuing time) goes to infinity, while if the packet queues are limited, or the network is a multi-drop network with many sources, and collisions may occur, the packet-dropping rate approaches 100%.

Line 31:

A well-known application of asymptotic throughput is in modeling [[point-to-point communication]] where (following Hockney) [[Network latency|message latency]] T(N) is modeled as a function of message length N as T(N) = (M + N)/A where A is the asymptotic bandwidth and M is the half-peak length.<ref>''Recent Advances in Parallel Virtual Machine and Message Passing Interface'' by Jack Dongarra, Emilio Luque and Tomas Margalef 1999 {{ISBN|3540665498}} page 134</ref>

As well as its use in general network modeling, asymptotic throughput is used in modeling performance on [[massively parallel]] computer systems, where system operation is highly dependent on communication overhead, as well as processor performance.<ref>M. Resch et al. ''A comparison of MPI performance on different MPPs''in Recent Advances in Parallel Virtual Machine and Message Passing Interface, Lecture Notes in Computer Science, 1997, Volume 1332/1997, 25-32</ref> In these applications, asymptotic throughput is used in Xu and Hwang model (more general than Hockney's approach) which includes the number of processors, so that both the latency and the asymptotic throughput are functions of the number of processors.<ref>''High-Performance Computing and Networking'' edited by Angelo Mañas, Bernardo Tafalla and Rou Rey Jay Pallones 1998 {{ISBN|3540644431}} page 935</ref>

===Peak measured throughput===

The above values are theoretical or calculated. Peak measured throughput is throughput measured by a real, implemented system, or a simulated system. The value is the throughput measured over a short period of time; mathematically, this is the limit taken concerningwith respect to throughput as time approaches zero. This term is synonymous with ''instantaneous throughput''. This number is useful for systems that rely on burst data transmission; however, for systems with a high [[duty cycle]], this is less likely to be a useful measure of system performance.

===Maximum sustained throughput===

This value is the throughput averaged or integrated over a long time (sometimes considered infinity). For high- duty cycle networks, this is likely to be the most accurate indicator of system performance. The maximum throughput is defined as the [[asymptotic throughput]] when the load (the amount of incoming data) is large. In [[packet switched]] systems where the load and the throughput always are equal (where [[packet loss]] does not occur), the maximum throughput may be defined as the minimum load in bit/s that causes the delivery time (the [[Network latency|latency]]) to become unstable and increase towards infinity. This value can also be used deceptively aboutin relation to peak measured throughput to conceal [[packet shaping]].

==Channel utilization and efficiency==

Throughput is sometimes normalized and measured in percentage, but normalization may confusecause confusion regarding what the percentage is related to. ''[[Channel utilization]]'', ''[[channel efficiency]]'' and ''[[packet drop rate]]'' in percentage are less ambiguous terms.

The channel efficiency, also known as [[bandwidth utilization efficiency]], is the percentage of the [[net bit rate]] (in bit/s) of a digital [[communication channel]] that goes to the actually achieved throughput. For example, if the throughput is 70&nbsp; Mbit/s in a 100&nbsp; Mbit/s Ethernet connection, the channel efficiency is 70%. In this example, effectively 70&nbsp; Mbit of data are transmitted every second.

Channel utilization is instead a term related to the use of the channel, disregarding the throughput. It counts not only with the data bits, but also with the overhead that makes use of the channel. The transmission overhead consists of preamble sequences, frame headers, and acknowledge packets. The definitions assume a noiseless channel. Otherwise, the throughput would not be only associated with the nature (efficiency) of the protocol, but also withto retransmissions resultingresultant from the quality of the channel. In a simplistic approach, channel efficiency can be equal to channel utilization assuming that acknowledge packets are zero-length and that the communications provider will not see any bandwidth relative to retransmissions or headers. Therefore, certain texts mark a difference between channel utilization and protocol efficiency.

In a point-to-point or [[point-to-multipoint communication]] link, where only one terminal is transmitting, the maximum throughput is often equivalent to or very near the physical data rate (the [[channel capacity]]), since the channel utilization can be almost 100% in such a network, except for a small inter-frame gap.

For example, the maximum frame size in Ethernet is 1526 bytes: up to 1500 bytes for the payload, eight bytes for the preamble, 14 bytes for the header, and 4 bytes for the trailer. An additional minimum inter-frameinterframe gap corresponding to 12 bytes is inserted after each frame. This corresponds to a maximum channel utilization of 1526&nbsp;/ (1526&nbsp;+ 12)&nbsp;× 100%&nbsp;= 99.22%, or a maximum channel use of 99.22&nbsp; Mbit/s inclusive of Ethernet datalink layer protocol overhead in a 100&nbsp; Mbit/s Ethernet connection. The maximum throughput or channel efficiency is then 1500&nbsp;/ (1526&nbsp;+ 12)&nbsp;= 97.5%, exclusive of the Ethernet protocol overhead.

==Factors affecting throughput==

Line 56:

The maximum achievable throughput (the channel capacity) is affected by the bandwidth in hertz and [[signal-to-noise ratio]] of the analog physical medium.

Despite the conceptual simplicity of digital information, all electrical signals traveling over wires are analog. The analog limitations of wires or wireless systems inevitably provide an upper bound on the amount of information that can be sent. The dominant equation here is the [[Shannon–Hartley theorem]], and analog limitations of this type can be understood as factors that affect either the analog bandwidth of a signal or as factors that affect the signal-to-noise ratio. The bandwidth of wired systems can be in fact surprisingly narrow, with the bandwidth of Ethernet wire limited to approximately 1&nbsp; GHz, and PCB traces limited by a similar amount.

Digital systems refer to the 'knee frequency',<ref>Johnson, 1993, 2-5</ref> the amount of time for the digital voltage to rise from 10% of a nominal digital '0' to a nominal digital '1' or vice versa. The knee frequency is related to the required bandwidth of a channel, and can be related to the [[3 db bandwidth]] of a system by the equation:<ref>Johnson, 1993, 9</ref> <math>\ F_{3dB} \approx K/T_r </math>

Where Tr is the 10% to 90% rise time, and K is a constant of proportionality related to the pulse shape, equal to 0.35 for an exponential rise, and 0.338 for a Gaussian rise.

*RC losses: Wires have an inherent resistance, and an inherent [[capacitance]] when measured concerningwith respect to ground. This leads to effects called [[parasitic capacitance]], causing all wires and cables to act as RC lowpass filters.

*[[Skin effect]]: As frequency increases, electric charges migrate to the edges of wires or cablescable. This reduces the effective cross-sectional area available for carrying current, increasing resistance and reducing the signal-to-noise ratio. For [[American wire gauge|AWG]] 24 wire (of the type commonly found in [[Cat 5e]] cable), the skin effect frequency becomes dominant over the inherent resistivity of the wire at 100&nbsp;kHz. At 1&nbsp; GHz the resistivity has increased to 0.1 ohm per inch.<ref>Johnson, 1993, 154</ref>

*Termination and ringing: Wires longer than about 1/6 wavelengths must be modeled as [[transmission line]]s with termination taken into account. Unless this is done, reflected signals will travel back and forth across the wire, positively or negatively interfering with the information-carrying signal.<ref>Johnson, 1993, 160-170</ref>

*[[Radio Propagation|Wireless Channel Effects]]: For wireless systems, all of the effects associated with wireless transmission limit the SNR and bandwidth of the received signal, and therefore the maximum bit [[transmission rate]].

===IC hardware considerations===

Computational systems have finite processing power and can drive finite current. Limited current drive capability can limit the effective signal- to- noise ratio for high [[capacitance]] links.

Large data loads that require processing impose data processing requirements on hardware (such as routers). For example, a gateway router supporting a populated [[class B subnet]], handling 10 × 100&nbsp; Mbit/s Ethernet channels, must examine 16 bits of address to determine the destination port for each packet. This translates into 81913 packets per second (assuming maximum data payload per packet) with a table of 2^16 addresses this requires the router to be able to perform 5.368 billion lookup operations per second. In a worst-case scenario, where the payloads of each Ethernet packet are reduced to 100 bytes, this number of operations per second jumps to 520 billion. This router would require a multi-teraflop processing core to be able to handle such a load.

* [[CSMA/CD]] and [[CSMA/CA]] "backoff" waiting time and frame retransmissions after detected collisions. This may occur in Ethernet bus networks and hub networks, as well as in wireless networks.

* [[flow control (data)|Flow control]], for example in the [[Transmission Control Protocol]] (TCP) protocol, affects the throughput if the [[bandwidth-delay product]] is larger than the TCP window, i.e., the buffer size. In that case, the sending computer must wait for acknowledgmentacknowledgement of the data packets before it can send more packets.

* TCP [[congestion avoidance]] controls the data rate. A so-called "slow start" occurs atin the beginning of a file transfer, and after packet drops caused by router congestion or bit errors in for example wireless links.

===Multi-user considerations===

Ensuring that multiple users can harmoniously share a single communications link requires some kind of equitable sharing of the link. If a bottleneck communication link offering data rate ''R'' is shared by "N" active users (with at least one data packet in the queue), every user typically achieves a throughput of approximately ''R/N'', if [[fair queuing]] [[best-effort]] communication is assumed.

* [[Packet loss]] due to [[network congestion]]. Packets may be dropped in switches and routers when the packet queues are full due to congestion.

* Packet loss due to [[bit error]]s.

* Scheduling algorithms in routers and switches. If fair queuing is not provided, users whothat send large packets will get higher bandwidth. Some users may be prioritized in a [[weighted fair queuing]] (WFQ) algorithm if differentiated or guaranteed [[quality of service]] (QoS) is provided.

* In some communications systems, such as satellite networks, only a finite number of channels may be available to a given user at a given time. Channels are assigned either through assignmentpreassignment or through Demand Assigned Multiple Access (DAMA).<ref>Roddy, 2001, 370 - 371</ref> In these cases, throughput is quantized per channel, and unused capacity on partially utilized channels is lost.

==Goodput and overhead==

{{main|Goodput}}

The maximum throughput is often an unreliable measurement of perceived bandwidth, for example, the file transmission data rate in bits per secondseconds. As pointed out above, the achieved throughput is often lower than the maximum throughput. Also, the protocol overhead affects the perceived bandwidth. The throughput is not a well-defined metric when it comes to how to deal with protocol overhead. It is typically measured at a reference point below the network layer and above the physical layer. The simplest definition is the number of bits per second that are physically delivered. A typical example where this definition is practiced is an Ethernet network. In this case, the maximum throughput is the [[gross bit rate]] or raw bit rate.

However, in schemes that include [[forward error correction codes]] (channel coding), the redundant error code is normally excluded from the throughput. An example is in [[modem]] communication, where the throughput typically is measured in the interface between the [[Point-to-Point Protocol]] (PPP) and the circuit-switched modem connection. In this case, the maximum throughput is often called [[net bit rate]] or useful bit rate.

To determine the actual data rate of a network or connection, the "[[goodput]]" measurement definition may be used. For example, in file transmission, the "goodput" corresponds to the file size (in bits) divided by the file transmission time. The "[[goodput]]" is the amount of useful information that is delivered per second to the [[application layer]] protocol. Dropped packets or packet retransmissions, as well as protocol overhead, are excluded. Because of that, the "goodput" is lower than the throughput. Technical factors that affect the difference are presented in the "[[goodput]]" article.

Line 95:

===Integrated circuits===

Often, a block in a [[data flow diagram]] has a single input and a single output, and operatesoperate on discrete packets of information. Examples of such blocks are [[Fast Fourier transform|Fast Fourier Transform]] modules or [[binary multiplier]]s. Because the units of throughput are the reciprocal of the unit for [[propagation delay]], which is 'seconds per message' or 'seconds per output', throughput can be used to relate a computational device performing a dedicated function such as an [[Application-specific integrated circuit|ASIC]] or [[embedded processor]] to a communications channel, simplifying system analysis.

===Wireless and cellular networks===