{
  "id": "erlang",
  "title": "Erlang",
  "lang": "en",
  "source_lang": "en",
  "summary": "The erlang is a dimensionless unit of traffic intensity used to measure resource occupancy and plan telecommunications capacity.",
  "category": "computer-science",
  "subcategory": "networks-security",
  "type": "concept",
  "aliases": [
    "erlang unit",
    "Erl",
    "traffic erlang"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/erlang/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/erlang/",
      "json": "https://aiwiki.page/zh/erlang.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "The erlang, symbol E, is a unit of traffic intensity used in telecommunications and other service systems. One erlang represents an average occupancy equivalent to one resource being continuously busy during a specified observation period. For example, 60 calls lasting one minute each, distributed over one hour, produce one erlang of carried traffic. The unit expresses resource demand or usage rather than the number of calls, the amount of transmitted information, or the duration of an individual connection.\n\nNamed after the Danish mathematician Agner Krarup Erlang, the erlang became central to the engineering of telephone exchanges and shared transmission facilities. It connects measurements of demand with probabilistic models that estimate how many circuits, agents, or other resources are required to meet a specified service objective. Its principal applications include circuit-group dimensioning, call-center staffing, and analysis of systems in which requests compete for a limited pool of service capacity.\n\nAlthough erlangs are particularly associated with circuit switching, their underlying idea—average simultaneous resource occupancy—applies more widely. Correct interpretation requires distinguishing traffic offered to a system from traffic actually carried, and recognizing that equal average loads can produce different blocking probabilities or waiting times."
    },
    {
      "heading": "Definition, dimensional meaning, and related quantities",
      "text": "Traffic intensity measures the average amount of a resource occupied over time. If \\(N(t)\\) is the number of resources busy at time \\(t\\), the carried traffic measured over an interval of duration \\(T\\) is\n\n\\[A_{\\mathrm{c}}=\\frac{1}{T}\\int_0^T N(t)\\,dt.\\]\n\nThe integral is the total occupied resource-time accumulated during the interval. Dividing it by the interval duration gives the average number of simultaneously occupied resources. If a group of ten circuits has an average of four circuits busy, it carries four erlangs. The traffic intensity is therefore dimensionless, although the named unit distinguishes it from other dimensionless quantities.\n\nFor a system with \\(n\\) identical resources, each capable of serving one request at a time, carried traffic cannot exceed \\(n\\) erlangs. The average utilization per resource is\n\n\\[u=\\frac{A_{\\mathrm{c}}}{n}.\\]\n\nThus, four erlangs carried by ten circuits correspond to an average utilization of \\(0.4\\), or 40 percent. This distinction is important: four erlangs describe aggregate occupancy, whereas 40 percent describes occupancy relative to the available capacity. An average of four busy circuits does not mean that exactly four are busy at every instant; the instantaneous number may fluctuate between zero and ten.\n\nTraffic volume is the accumulated resource-time rather than its average rate. Over an interval \\(T\\), it is\n\n\\[V=A_{\\mathrm{c}}T.\\]\n\nIt may be expressed in erlang-hours or erlang-seconds. Two erlangs sustained for three hours correspond to six erlang-hours. Another traditional measure is the hundred-call-second, commonly abbreviated CCS. One erlang sustained for one hour represents 3,600 call-seconds, or 36 CCS. Converting a CCS count into erlangs therefore requires knowing the measurement interval.\n\nErlangs are not a measure of data-transfer rate. A voice circuit transmitting at a fixed bit rate can carry one call continuously and therefore contribute one erlang, regardless of how much meaningful speech occurs. Conversely, two services occupying the same number of circuits can have different coding rates and transmit different numbers of bits. Converting erlangs into bandwidth requires an additional model specifying the capacity consumed by each occupied resource.\n\nThe identity of the resource must consequently be explicit. An erlang may refer to occupancy of a telephone circuit, a radio traffic channel, a server, or an agent position. A session that occupies several resources simultaneously contributes separately to each relevant resource pool; its traffic cannot be interpreted without identifying which pool is being measured."
    },
    {
      "heading": "Historical development and standardization",
      "text": "The erlang emerged from the practical problem of sharing expensive telephone facilities. Early exchanges could not economically provide a dedicated interexchange circuit for every subscriber pair. Instead, many potential callers shared smaller groups of trunks. Engineers needed a way to determine how large those groups should be, given the frequency and duration of calls and an acceptable probability that no circuit would be available.\n\nAgner Krarup Erlang, who worked for the Copenhagen Telephone Company, developed a mathematical treatment of these questions in the early twentieth century. His 1909 paper on probability and telephone conversations was an important early contribution to the statistical analysis of telephone demand. In 1917 he published work containing results associated with the classical loss and waiting-time formulas now known as Erlang B and Erlang C. These results connected random call arrivals, call durations, and finite service capacity.\n\nThe essential departure from simple arithmetic was to treat occupancy as a fluctuating quantity. If an exchange receives an average demand of eight simultaneous calls, eight circuits do not necessarily provide adequate service. There will be periods with fewer than eight calls and periods when demand exceeds eight. The engineering question is therefore not merely the average load, but the probability and consequences of exceeding available capacity.\n\nOther researchers developed related approaches. The British engineer Edward C. Molina contributed to the mathematical theory of telephone traffic, while Tore Engset developed a finite-source loss model. The Engset formula addresses situations in which the number of potential callers is limited and the arrival rate depends on how many sources are already busy. These developments helped establish teletraffic engineering as a quantitative discipline rather than a collection of local sizing rules.\n\nThe International Telephone and Telegraph Consultative Committee, generally known by its French initials CCITT, adopted the name erlang for the traffic unit in 1946. The symbol E and the associated traffic concepts became part of international telecommunications terminology. Naming the unit after Erlang also reflected the close connection between the measurable quantity and the probabilistic dimensioning methods used with it.\n\nThe historical distinction between unit and formula remains important. An erlang is a measure of traffic intensity; Erlang B and Erlang C are particular mathematical models. A traffic measurement can be expressed in erlangs even when neither model is appropriate. Likewise, the programming language Erlang is a separate subject: it is not a notation or software implementation required for measuring telecommunications traffic."
    },
    {
      "heading": "Offered traffic, carried traffic, and measurement",
      "text": "The most important operational distinction is between offered traffic, representing demand presented to a resource group, and carried traffic, representing demand actually served. In a conventional model, offered traffic is\n\n\\[A=\\lambda h,\\]\n\nwhere \\(\\lambda\\) is the mean rate of arriving requests and \\(h\\) is their mean service or holding time. Both quantities must use compatible time units. If 120 calls arrive per hour and occupy a circuit for an average of three minutes, the offered traffic is\n\n\\[A=120\\times\\frac{3}{60}=6\\ \\mathrm{E}.\\]\n\nThis calculation assumes that the duration assigned to an offered request represents the service it would require if admitted.\n\nCarried traffic is usually measured from actual occupancy records. If accepted calls occupy circuits for a total of 18,000 seconds during a one-hour interval, the group carries five erlangs. A system that blocks calls can have offered traffic greater than carried traffic. Under the standard loss-model assumptions, if the call-blocking probability is \\(B\\), the relationship is\n\n\\[A_{\\mathrm{c}}=A(1-B).\\]\n\nThe corresponding lost traffic is \\(AB\\). Outside those assumptions, the fraction of calls blocked and the fraction of offered workload lost need not be identical, particularly when acceptance depends on request size or class.\n\nObservation boundaries require care. A call beginning before the measurement interval still contributes occupancy during the interval, and a call ending afterward contributes only the part falling within it. Counting only calls that both begin and end inside the window can bias an estimate, especially when the window is short relative to holding times. Direct integration of busy-resource counts avoids this particular boundary problem.\n\nDemand also varies systematically with time. Telephone networks have historically been dimensioned using busy-hour measurements, but the busiest hour may differ across locations, weekdays, seasons, or service classes. A daily average can conceal short periods of severe congestion. An average of ten erlangs over a full day does not imply that a ten-circuit group can accommodate the peak hour.\n\nMeasurements must specify what constitutes resource occupation. Depending on the network, a circuit can be held during call setup, ringing, conversation, and release. At a contact center, an agent’s service time may include talk time, time spent placing the caller on hold, and after-call work. These definitions materially affect the calculated load.\n\nRetries create a further complication. A blocked caller may dial again, increasing the number of attempts observed by the system without creating an equal number of independent underlying communication needs. Measurements that mix original attempts and retrials can therefore misrepresent the demand assumed by a simple arrival model."
    },
    {
      "heading": "Stochastic foundations and the meaning of congestion",
      "text": "An average traffic intensity alone does not determine service performance. A resource group offered ten erlangs may perform well if demand is smooth, but poorly if requests arrive in concentrated bursts. Classical traffic theory supplements the mean load with assumptions about arrival processes, service times, resource sharing, and the treatment of requests that encounter a full system.\n\nA central model assumes a Poisson arrival process. With a constant arrival rate \\(\\lambda\\), the number of arrivals in an interval of duration \\(t\\) has a Poisson distribution with mean \\(\\lambda t\\). Counts in disjoint intervals are independent. This model can approximate the aggregate behavior of many independent callers, each making requests relatively infrequently, but it does not automatically describe synchronized events or repeated attempts following congestion.\n\nHolding times are another component. An exponential distribution with mean \\(h\\) has the memoryless property: the conditional distribution of the remaining duration does not depend on how long service has already lasted. Combining Poisson arrivals with exponential service times often produces a tractable Markov chain, in which the number of requests present is sufficient to describe the system’s probabilistic evolution.\n\nAn unlimited-capacity system provides a useful reference. With Poisson arrivals and independent holding times, the stationary number of active requests is Poisson distributed with mean \\(A=\\lambda h\\). Its variance is also \\(A\\). A finite resource group modifies this occupancy process because excess demand must either be rejected, queued, or redirected. The choice among these mechanisms leads to different dimensioning formulas.\n\nCongestion itself has several meanings. Time congestion is the proportion of time during which every resource is busy. Call congestion is the proportion of arriving requests that find the resources unavailable. Under Poisson arrivals, the property called Poisson arrivals see time averages, or PASTA, makes these probabilities equal in the standard setting. They need not be equal when arrivals depend on system occupancy or on a finite population of sources.\n\nAnother broad result is Little’s law,\n\n\\[L=\\lambda W,\\]\n\nwhere \\(L\\) is the long-run average number of customers in a defined system, \\(\\lambda\\) is the throughput through that system, and \\(W\\) is average time spent there. Unlike a specific Erlang formula, this relation does not require exponential service times. Its variables must refer to the same system boundary and customer population. In a stable queue, customers waiting and customers receiving service both contribute to \\(L\\); resource occupancy alone counts only those in service."
    },
    {
      "heading": "Erlang B and systems that reject excess demand",
      "text": "The Erlang B formula describes a fully accessible group of \\(n\\) identical resources in which an arrival is immediately rejected if all resources are occupied. Rejected requests do not wait within the model. Under Poisson arrivals, independent identically distributed holding times, and no state-dependent change in demand, its blocking probability is\n\n\\[B(n,A)= \\frac{A^n/n!} {\\displaystyle\\sum_{k=0}^{n}A^k/k!},\\]\n\nwhere \\(A\\) is the offered traffic in erlangs. Typical applications include telephone trunk groups and other admission-controlled systems with a fixed number of simultaneous service positions.\n\nThe formula can be understood through the stationary occupancy probabilities. For \\(k=0,\\ldots,n\\),\n\n\\[p_k= \\frac{A^k/k!} {\\displaystyle\\sum_{j=0}^{n}A^j/j!}.\\]\n\nBlocking occurs when an arrival finds the system in state \\(n\\). Poisson arrivals see the time-average occupancy distribution, so \\(B(n,A)=p_n\\). The mean carried traffic is \\(A[1-B(n,A)]\\), and the utilization of each resource in a symmetric group is this quantity divided by \\(n\\).\n\nAn important property is insensitivity to the detailed holding-time distribution. Although a common derivation uses exponential holding times, the stationary blocking result remains valid for independent identically distributed holding times with the same finite mean under the standard Poisson loss-system assumptions. Thus, a nonexponential duration distribution does not by itself invalidate Erlang B. Correlation, state-dependent arrivals, nonuniform resource requirements, and retrials can still invalidate the simple model.\n\nFor example, ten circuits offered eight erlangs have a blocking probability of approximately \\(0.1217\\), or 12.17 percent. The carried traffic is approximately\n\n\\[8(1-0.1217)\\approx7.03\\ \\mathrm{E}.\\]\n\nThe group therefore has an average utilization of about 70.3 percent while still rejecting more than one call in ten. This illustrates why unused average capacity does not imply an absence of congestion: available circuits at quiet moments cannot be stored for use during peaks.\n\nDirect evaluation with factorials and high powers can become numerically inconvenient. A stable recursive computation begins with\n\n\\[B(0,A)=1\\]\n\nand proceeds through\n\n\\[B(n,A)=\\frac{A B(n-1,A)}{n+A B(n-1,A)}.\\]\n\nThis permits efficient evaluation using an iterative algorithm and avoids forming very large intermediate terms.\n\nThe formula represents complete sharing: every arriving call can use any free resource in the group. Restricted routing, reserved circuits, priorities, or calls that require several circuits need other models. Similarly, overflow traffic arriving from an already congested group is generally more variable than a fresh Poisson stream, so treating it as ordinary offered traffic can underestimate blocking."
    },
    {
      "heading": "Erlang C and systems in which requests wait",
      "text": "The Erlang C formula models a system with \\(n\\) identical servers and a waiting room of unlimited capacity. Arrivals finding every server busy join a queue rather than being lost. The classical model assumes Poisson arrivals, independent exponential service times, and a common first-come, first-served queue. It is conventionally described as an \\(M/M/n\\) queue, with the two letters indicating Markovian arrival and service assumptions.\n\nLet the mean service time be \\(h\\), so that each server’s service rate is \\(\\mu=1/h\\), and let \\(A=\\lambda h\\). A stationary distribution exists only when\n\n\\[A<n.\\]\n\nEquivalently, utilization \\(\\rho=A/n\\) must be less than one. The probability that an arriving customer must wait is\n\n\\[C(n,A)= \\frac{ \\displaystyle\\frac{A^n}{n!}\\frac{n}{n-A} }{ \\displaystyle\\sum_{k=0}^{n-1}\\frac{A^k}{k!} + \\displaystyle\\frac{A^n}{n!}\\frac{n}{n-A} }.\\]\n\nUnlike Erlang B, this is a delay probability, not a probability of rejection.\n\nThe mean queueing delay is\n\n\\[\\mathbb{E}[W_q]=\\frac{C(n,A)}{n\\mu-\\lambda} =\\frac{C(n,A)h}{n-A}.\\]\n\nFor a waiting-time threshold \\(t\\geq0\\), the probability of waiting longer than \\(t\\) is\n\n\\[P(W_q>t)=C(n,A)e^{-(n\\mu-\\lambda)t}.\\]\n\nThese expressions separate the probability of any wait from the duration of waiting. They also show the sharp deterioration as offered traffic approaches total service capacity: the denominator \\(n\\mu-\\lambda\\) becomes small, and mean delay grows rapidly.\n\nFor ten servers offered eight erlangs, the delay probability is approximately 0.4092. If mean service time is three minutes, the mean queueing delay is about 36.8 seconds. The probability of beginning service within 20 seconds is approximately\n\n\\[1-0.4092e^{-2(20/180)} \\approx0.672,\\]\n\nor 67.2 percent. The servers average 80 percent utilization, since a stable system without abandonment eventually serves every arrival.\n\nErlang C has been widely used in contact-center planning, where the servers are agents and the holding time is average handling time. Its assumptions nevertheless matter. Actual callers may abandon, agents may handle different skill groups, and service times may be far from exponential. Unlimited waiting-room capacity is also an idealization.\n\nFor a fixed load and server count, Erlang B and Erlang C answer different questions and cannot be substituted for one another. In a loss system, rejected requests disappear immediately; in a waiting system, deferred work remains and affects subsequent arrivals. At or above \\(A=n\\), Erlang C has no finite stationary mean queue, whereas an Erlang B system still has a stationary occupancy distribution because excess demand is discarded."
    },
    {
      "heading": "Capacity planning, pooling, and service objectives",
      "text": "Dimensioning reverses the calculation performed by a traffic model. Rather than asking how much congestion a given resource group experiences, it asks how many resources are needed for a forecast load and a specified performance target. In a loss system, the target may be a maximum blocking probability. In a queueing system, it may be the fraction of requests beginning service within a threshold, or a limit on mean waiting time.\n\nFor a blocking target \\(b\\), the required capacity is the smallest integer \\(n\\) satisfying\n\n\\[B(n,A)\\leq b.\\]\n\nThe integer constraint matters because capacity is added in discrete circuits, channels, or service positions. With ten erlangs of offered traffic, Erlang B gives blocking of about 1.09 percent for 17 circuits and about 0.71 percent for 18 circuits. A strict 1 percent target therefore requires 18 circuits under the model, not merely ten resources to match the mean load.\n\nSharing capacity produces trunking efficiency, also called a pooling gain. A large common group can usually carry more traffic at a given blocking probability than several smaller isolated groups with the same total number of resources. Independent fluctuations partly offset one another, allowing a free resource to serve demand that would otherwise be stranded in another group. The gain depends on actual sharing: adding capacities on paper does not create pooling if calls cannot access the combined resources.\n\nFor large many-server systems, square-root staffing describes an important scaling principle. Capacity is often approximated by\n\n\\[n\\approx A+\\beta\\sqrt{A},\\]\n\nwhere the safety factor \\(\\beta\\) depends on the service objective and model. The mean demand grows proportionally to \\(A\\), while a statistical safety allowance grows on the order of \\(\\sqrt{A}\\). This helps explain why larger pools can achieve both high utilization and relatively good service, although the approximation is not a universal replacement for exact calculations.\n\nOperational capacity can differ from installed or scheduled capacity. A telephone group may lose circuits to failures or maintenance. A contact center may have scheduled agents who are unavailable because of breaks, training, or other duties. Calculating the required number of simultaneously available agents is therefore separate from determining the total roster needed to provide them. Treating these quantities as identical understates staffing requirements.\n\nForecast uncertainty is distinct from random variation within a traffic model. Erlang calculations condition on an assumed arrival rate and mean holding time; they do not automatically account for errors in either input. If both increase, their effects multiply through \\(A=\\lambda h\\). Service objectives may also be defined over different intervals or customer classes, so an acceptable daily average can coexist with poor performance during the busiest period."
    },
    {
      "heading": "Extensions beyond the classical models",
      "text": "Classical Erlang models impose a particular relationship between demand and available resources. Several extensions alter that relationship to represent systems more closely. One major distinction is between an effectively infinite population of potential requesters and a finite source population. If only a small number of devices can generate requests, a device already in service cannot simultaneously generate another request of the same kind.\n\nThe Engset model accounts for this finite-source effect. Its total arrival rate depends on the number of idle sources and decreases as more sources become busy. The blocking probability experienced by attempts must be distinguished from the fraction of time all resources are occupied. Erlang B becomes a useful approximation when the source population is large and each source individually contributes a small amount of traffic.\n\nAbandonment requires a different extension. In the Erlang A model, the classical multiserver queue includes exponentially distributed customer patience times. A customer may leave before service begins. This can stabilize the queue even when offered workload exceeds nominal service capacity, but stability does not mean adequate service: a large share of customers may abandon. Arrival rate multiplied by desired service duration describes potential workload, while actual server occupancy depends on how many requests survive to service.\n\nRetrial systems represent customers who return after rejection or an unsuccessful attempt. A pool of retrying customers introduces dependence between past congestion and future arrivals. Repeated attempts can amplify a demand peak and invalidate the assumption that the observed stream consists of independent Poisson arrivals. Models may include an orbit of waiting retrial customers, explicit retry delays, or approximations based on measured repeat-attempt behavior.\n\nMultirate loss systems arise when requests consume different amounts of a shared capacity. One connection might require one capacity unit while another requires several. A single erlang total then loses important information: one erlang of narrowband calls and one erlang of wideband calls do not impose the same capacity requirement. Loss-network models retain traffic intensity for each class together with its resource requirement; methods such as Kaufman–Roberts recursion calculate blocking in important single-resource multirate cases.\n\nNetworks of resource groups introduce routing and dependence. A connection may require simultaneous capacity on several links, or may be sent to an alternative route when its preferred route is full. Blocking on one group changes the traffic offered elsewhere. Priorities, reservations, and skill-based assignment add further structure. In these settings, analytical approximations and discrete-event simulation are used to evaluate measures that cannot be determined from one aggregate erlang figure alone."
    },
    {
      "heading": "Packet networks, computing systems, and limits of interpretation",
      "text": "In packet-switched networks, a communication session does not necessarily occupy a fixed transmission channel for its entire duration. Packets from many sessions share links through multiplexing, and the instantaneous transmission demand can vary sharply within each session. Consequently, an erlang measurement of active sessions is not sufficient to dimension a packet link without assumptions about packet rates, packet sizes, activity patterns, and queueing behavior.\n\nA link’s average workload can still be written in a form resembling traffic intensity. If packets arrive at rate \\(\\lambda_p\\), have mean length \\(\\mathbb{E}[L]\\) bits, and traverse a link of rate \\(R\\) bits per second, the average transmission load is\n\n\\[\\rho=\\frac{\\lambda_p\\mathbb{E}[L]}{R}.\\]\n\nThis ratio is dimensionless and describes the fraction of link capacity demanded on average under the stated model. It is usually called utilization or load rather than being reported in erlangs. Even where average utilization is moderate, bursts can fill buffers and produce delay or packet loss.\n\nVoice services carried over the Internet illustrate the distinction between session occupancy and transmission occupancy. An active call may remain established throughout a conversation, while voice activity detection reduces the packets sent during silence. Codec choice, packet headers, redundancy, and packetization intervals affect bandwidth. Thus, a count of simultaneous voice calls can support an erlang-style session model, but bandwidth estimation requires these additional technical parameters.\n\nSimilar reasoning applies to computing infrastructure. Requests entering a pool of workers generate an average workload equal to arrival rate multiplied by mean time actually occupying a worker. That quantity can be interpreted as demand measured in server equivalents. However, request elapsed time may include waiting for storage, network responses, or locks; it is not automatically equal to CPU service time. An asynchronous service may keep many requests in progress while only a few consume processor time at once.\n\nIn cloud computing and load balancing, capacity may change during operation, service instances may have unequal performance, and request classes may compete for several resources. A request can be CPU-bound, memory-bound, or constrained by a database connection pool. Aggregating all demand into one erlang value obscures these bottlenecks. Models often need a separate service demand and capacity constraint for each resource.\n\nThe continuing usefulness of the erlang lies in its clearly defined measurement boundary. It describes how much occupancy or service demand is associated with a specified resource pool over a specified interval. Predictions of blocking, delay, or required capacity additionally depend on arrival variability, service-time behavior, routing rules, and the treatment of unserved requests. Reporting those assumptions alongside the traffic intensity preserves the distinction between a measured load and the performance model used to interpret it."
    }
  ],
  "links": [
    "telephone",
    "circuit-switching",
    "bit",
    "programming-language",
    "poisson-process",
    "exponential-distribution",
    "markov-chain",
    "variance",
    "algorithm",
    "packet-switching",
    "multiplexing",
    "internet",
    "central-processing-unit",
    "cloud-computing",
    "load-balancing",
    "database"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}