Congestion window 2026-10-31
A congestion window bounds a sender's outstanding data according to congestion control. In TCP it is local sender state, distinct from the advertised receive window.
Elena Yudovina 2026-11-05
Elena Yudovina is a mathematician whose work includes the theory of stochastic networks. With Frank Kelly, she wrote Stochastic Networks by Frank Kelly and Elena Yudovina, connecting queueing theory, loss networks, congestion control and distributed optimization. The author's book page verifies the joint authorship and 2014 publication.[ref]
A central question is how local service or routing rules affect an entire network. The book's M/M/1 queue illustrates the danger of operating near capacity: with arrival rate , service rate and , mean queue population is . Modest changes in load can therefore cause large changes in delay.[ref]
Frank Kelly 2026-11-05
Frank Kelly works on probability theory, stochastic networks and the mathematical organization of shared resources. His research connects queueing theory with optimization: local decisions about service, admission or price can determine a network's aggregate performance. His institutional research record describes these interests.[ref]
With Elena Yudovina, Kelly wrote Stochastic Networks by Frank Kelly and Elena Yudovina. Its treatment of loss networks shows how detailed balance can produce an explicit equilibrium occupancy law, while congestion control uses resource prices to coordinate competing demands. These are complementary uses of mathematics: predicting congestion and designing rules that improve it.[ref]
UDP fits IP multicast because each independent datagram can be delivered to a group without establishing a separate transport connection with every receiver. Ordinary TCP defines a reliable ordered byte stream between two endpoints, with one peer's acknowledgements, receive window and connection state. A group has many receivers with different losses and capacities, so that interface cannot simply be assigned to one group address. A reliable group service can be designed, but ordinary TCP does not provide it. UDP itself adds neither group reliability nor congestion control; applications must provide whatever additional behaviour they require.
A virtual multicast distribution tree lets one packet traverse a shared upstream link and be copied only where branches diverge. Compared with a separate unicast copy per receiver, this saves repeated transmission over the common links. A tree also supplies a unique parent route and avoids forwarding loops when the routing information is consistent. Branches with no interested receivers can be pruned.
For a source-specific multicast distribution tree, a router needs the group identifier and source identifier, the expected incoming interface, the set of outgoing interfaces leading to interested receivers, and information about local group membership. Timers and join/prune information maintain those decisions. This is naturally keyed by for source and group ; other multicast routing designs can use a shared group tree. A router needs the relevant local forwarding entries, rather than a complete list of all group members everywhere. The result is
Flow control prevents a sender from overrunning the receiving endpoint's capacity to buffer and consume data. Congestion control prevents the aggregate traffic from overloading the intervening computer network, especially shared links and router queues. A fast receiver can still sit beyond a congested path; an uncongested path can still lead to a slow receiver. Consequently
TCP implements both restrictions at the sending endpoint, using information from different places. For flow control, the receiving TCP advertises its available receive window in TCP headers. The sending TCP tracks that advertised range together with the acknowledgement number, and must not send ordinary new data beyond its right edge. As the receiving application consumes data, the receiver can advertise more space. A zero receive window stops ordinary new data and invokes the separate probing mechanism.
For congestion control, the sender maintains a congestion window representing its permitted amount of outstanding data given the path's condition. Acknowledgements, duplicate acknowledgements and data retransmission timeouts supply feedback; explicit congestion notifications can also contribute when negotiated. Routers supply the underlying evidence through queuing, drops or marks, but ordinary TCP congestion window management takes place at the endpoints, rather than through a per-connection capacity reservation at each router.[ref]
Ignoring exceptional probe behaviour, let be the number of sent but not yet acknowledged bytes. The two controls combine approximately as
The send window is anchored by the acknowledged sequence position, rather than authorizing a fresh windowful on every send. Acknowledgements release room as old data leave flight. A sudden reduction of a limit can leave already transmitted data above it; the sender then waits instead of recalling those bytes. The receive window is explicit receiver feedback, whereas the congestion window is a local sender variable. No details of their parameter-adjustment algorithms are needed for this distinction.
For two competing TCP connections, each sender's congestion control aims to obtain useful throughput while adapting to shared capacity and avoiding persistent overload. For two comparable, continuously backlogged connections at one bottleneck, the usual fairness objective is approximately half the available capacity each, while keeping the bottleneck productively used. The goal is not for both senders independently to seize the whole link. Unequal delays, implementations, application demands and other bottlenecks mean that equal throughput is not guaranteed.
For two UDP exchanges, there is no corresponding built-in transport congestion control objective: UDP sends datagrams without specifying a congestion response. The applications might implement well-behaved congestion control, rate limiting or no adaptation at all. Consequently two UDP labels alone do not determine either a target share or the resulting allocation. Applications should still manage overload; that responsibility is outside the base UDP service.
For a competing TCP connection and an unresponsive high-rate UDP application sharing an ordinary queue, losses cause the TCP sender to reduce its offered load while the UDP sender continues. The UDP traffic can take much of the bottleneck capacity and severely reduce TCP throughput. If the UDP rate exceeds capacity, it too loses packets; sending more does not guarantee useful delivery. A modest-rate or adaptive UDP application, or a router enforcing fair queueing, gives a different outcome. Thus
By January 2020, the Internet had retained a broadly interoperable packet-delivery substrate while acquiring substantial centralisation of services and more functionality inside communication paths. The departure is real but uneven: independent network operation, physical replication, application control and placement of security functions are different architectural dimensions. Future service design should preserve endpoint responsibility for application correctness and make any delegation of control, state or trust explicit.
Network decentralisation means that important decisions and dependencies are spread across independently controlled participants, rather than being concentrated in one indispensable operator. It is not established merely by counting machines. A distributed system can run in many locations under one administrative authority. The end-to-end principle addresses a different question: where should a function be implemented? If fully satisfying an application requirement needs knowledge available at the communicating applications, an implementation confined to the intervening network cannot replace that application-level responsibility. The two ideas often reinforce one another, because a general communication service leaves applications free to evolve, but they are not equivalent.
The original argument by Jerome H. Saltzer, David P. Reed and David D. Clark is about completeness of a system function, not a blanket prohibition on intelligent routers. Consider transferring a file whose correctness includes the bytes eventually stored at the destination. Perfect link checks do not exclude a subsequent storage error. The application still needs a check at the relevant endpoint. Local error recovery may improve efficiency, while the end-to-end check establishes the requested property over the complete path through the system. The reusable argument is that application correctness is not determined by a network trace whenever two executions can have identical network behaviour but different endpoint outcomes.[ref]
This division also explains the Internet's layered network architecture. Internet Protocol provides a common packet-delivery interface across different underlying networks; it does not promise reliable, ordered application completion. Transmission Control Protocol implements a reliable ordered byte stream between its transport endpoints, but receiving bytes is still different from completing the application's intended operation. Fate-sharing places essential communication state with the hosts that depend on it, reducing the need to preserve that state in every intermediate forwarding device. A route change can then leave the endpoint state intact, provided connectivity returns and the protocol can recover before it gives up. This does not mean routers contain no state: forwarding tables are different from indispensable per-connection state.[ref]
The IP hourglass gives this separation an evolutionary advantage. Many applications can use the same narrow inter-network interface, while many transmission technologies can support it. An application need not redesign every intermediate network before it can introduce a new service. However, the early architecture was not an entirely flat collection of equal peers. The Domain Name System is both hierarchical and distributed, with delegated authority and replicated information. Its design replaced a centrally distributed host-name file, so some architectural developments increased distribution. A common protocol or a logically coordinated namespace is compatible with independently administered components. The relevant baseline is a set of design goals and compromises, not an earlier Internet with no servers, hierarchy or coordination.[ref][ref]
A clear departure from direct endpoint reachability is network address translation, particularly network address and port translation. A translator can let several privately addressed hosts share an externally usable address by maintaining mappings for their traffic. This helped accommodate address scarcity and could be introduced incrementally. It also made the path depend on an additional stateful intermediary: incoming traffic generally needs an appropriate mapping or explicit configuration, mappings can expire, and moving a flow to a different translator can break it unless the necessary state is available there. Two applications being willing to communicate is consequently not sufficient for connectivity. Application-specific gateways and relays may become necessary. The benefit is practical deployment; the cost is reduced address transparency and extra recovery and deployment constraints.[ref]
Firewalls and application proxies make another important qualification to the ideal of an unrestricted path. A firewall deliberately restricts traffic according to a security policy. An application proxy can perform a function on behalf of an endpoint, sometimes with information and permissions beyond ordinary packet forwarding. Such arrangements can protect vulnerable systems or provide a useful service; treating every intervention as an architectural mistake would ignore actual security and operational requirements. The cost must nevertheless be recognised: an intermediary can become an additional failure point, a source of incompatible assumptions or an authority whose permission is needed to deploy a new application. Different endpoint operators may have conflicting objectives, so assuming universally cooperative, trustworthy hosts is insufficient as a design method.[ref][ref]
One consequence is protocol ossification. If devices along a path recognise only familiar packet formats or alter fields in undocumented ways, changing the transport endpoints alone may not be enough to deploy a new transport protocol. An extension can fail even when the two hosts support it. This is a loss of the original freedom to innovate at the edges, rather than simply an increase in the number of network devices. By 2015 an architectural workshop had explicitly examined middlebox and interface constraints on transport evolution. A new service therefore needs a deployment strategy that accounts for real paths, including appropriate negotiation, measurement and fallback, rather than assuming every syntactically valid packet will pass.[ref]
Application architecture has also become less decentralised in a way that is not reducible to middleboxes. A service can concentrate user accounts, authoritative data and protocol decisions in a provider while running many geographically separated servers. Cloud computing and a content delivery network can reduce the effort needed to deploy a service, improve network latency by placing content closer to users, and spread load across sites. A content delivery network can direct requests to delivery servers and use caching to keep copies near demand; the origin provider can authorise this arrangement. These are substantial engineering gains. Yet losing access to a provider's identity, configuration or authoritative data service can disable an application even while IP forwarding remains available. Geographical replication does not imply decentralised control.[ref]
A browser communicating with a central application server is not, merely for being client-server model, a violation of the end-to-end principle. The server may be exactly the application endpoint that has the necessary knowledge. The architectural concern is whether users can choose or replace that endpoint, whether independent implementations can interoperate, and what additional parties must be trusted. Internet service federation addresses the administrative dimension: independent providers communicate using an agreed protocol instead of every participant needing an account at one authority. Internet mail supplies an established example of communication across separately operated domains, even though individual organisations may use large central servers. Federation does not itself guarantee cryptographic confidentiality, fault tolerance or choice among many operators; it changes who can participate and which dependencies are mandatory.[ref]
The distinction is especially clear for encryption. Transport Layer Security protects a channel between its cryptographic endpoints; it does not define those endpoints as the two ultimate human users. In one design, a browser's TLS peer is the origin server and intermediate routers lack the channel keys. In another, the origin authorises a CDN edge to terminate TLS and the CDN edge uses a second protected channel to the origin. Both links can be encrypted while the CDN edge processes plaintext. TLS termination moves the confidentiality boundary. This is not automatically an unauthorised interception, but it is a different trust arrangement that should be documented. Network address translation alone does not require decrypting a TLS payload.[ref]
Figure 1.
Forwarding and address translation differ from authorised TLS termination: two protected channels can expose plaintext at their shared endpoint
. Details.
If a service promises secrecy from its own relay or storage provider, transport encryption to that provider is insufficient. End-to-end encryption must keep the relevant plaintext and decryption capability at the intended application endpoints. Similarly, a client can verify a publisher's digital signature on an object obtained through a cache, separating the source of delivery from the authority for content. Neither mechanism alone hides all communication metadata, solves endpoint compromise, or ensures the service is reachable. The design question is which party needs which information, and what property the user is actually promised, rather than whether every segment has an encryption label.
These changes therefore call for selective preservation of the original principles. At the network layer, favour general interfaces and independently deployable extensions rather than unnecessary application-specific assumptions. IPv6 expands the address space and can remove a reason for address sharing; it does not compel operators to permit incoming connections, remove firewalls or decentralise application ownership. It would be unjustified to expect a larger address alone to restore the whole earlier architectural ideal. Address transparency, security policy and application control require separate choices.[ref][ref]
At the transport layer, the November 2019 QUIC Internet-Draft provides a timely example of trying to evolve functionality at the endpoints over the widely deployed User Datagram Protocol. It includes encrypted transport machinery, connection identifiers and support for changes of network path. As of this examination it is work in progress, not a completed Internet standard. Its approach shows how deployment constraints can stimulate a new endpoint implementation, but it is not proof that every network will carry UDP or that encapsulation abolishes all middlebox dependencies. Designers must preserve congestion control and explicitly assess the tradeoff between encrypted protocol details and information needed for network operation.[ref]
At the application layer, put the final checks for the promised result at the endpoints that can actually observe it. An intermediate acknowledgement should not be confused with successful storage or completion of a user operation. When an intermediary is useful, specify its authority, its failure behaviour and the state it holds; make security-sensitive delegation visible rather than silently expanding trust. Separate this from the performance decision to use data replication or caching. Those optimisations may be valuable even when endpoint verification remains indispensable.
Resilience likewise requires examining dependencies across operators and components, not just counting replicas. Identify which failures can be tolerated while an application still makes progress, including failure of a naming, identity or configuration service. Where appropriate, use independently operated services, interoperable protocols and data portability so that users can move state or change providers. These options have costs: federation requires compatibility between implementations, while replicas need coherent handling of updates. A centrally operated component can be a reasonable choice when its benefits justify a clearly identified dependency. It should not acquire the label decentralised simply because it is replicated.
A useful evaluation would test the service from independently administered networks, including one with address translation, and distinguish connection establishment, endpoint correctness and recovery after an intermediary restart. For a delegated delivery service, separately test what an authorised CDN edge can observe and what an application-level authenticity check validates. Such tests make the architectural claims concrete; a successful throughput measurement alone cannot establish freedom from mandatory intermediaries or secrecy from a delivery provider.
The strongest January 2020 assessment is that end-to-end function placement remains a valuable criterion, while universal reachability and independent control have been constrained at several layers. The design goal for subsequent services is a clear allocation of correctness, trust and failure responsibility, with the benefits and dependencies of intermediaries exposed rather than hidden behind a simple centralised/decentralised label.
Each user maintains a transmission rate . Resource measures its aggregate load and returns the resource congestion price . The user adds these feedback signals to form its route congestion price . Its increase term is the constant , while its decrease term is proportional to both its current rate and the congestion signal. This is a distributed congestion control algorithm: resources need only local loads, and users need only prices along their own routes.
There is a needed distinction in the word increasing. If it permits nondecreasing price functions, the all-zero prices are allowed. Then , so there is no equilibrium. The intended convergence statement holds when each nonempty route meets at least one resource whose price is positive at some load. This is automatic if every price is strictly increasing and nonnegative. We prove it under the more general nondecreasing assumption with this explicit condition, for finite sets of nonempty routes and nonnegative initial rates.
Write . Each is convex and nondecreasing, and the primal congestion potential is
On positive rates this is strictly concave: the sum of positive-weight logarithms is strictly concave, while the resource term is convex. Its derivatives are
Consequently the gradient flow with diagonal mobility satisfies
with equality exactly at an equilibrium.
We must also prove trapping and existence of the equilibrium. For each route choose a resource on it and constants such that its price is at least beyond load . Finiteness gives common constants . If is large, a resource on a maximizing route has load at least , so the resource penalty is at least . With , therefore,
Once all rates are bounded above, a rate tending to zero also makes through its positive-weight logarithm. Thus every nonempty superlevel set of is compact inside the positive orthant. A maximum exists by restricting to such a superlevel set, and strict concavity makes it unique, say . It is interior and satisfies , equivalently
Conversely every positive equilibrium is a stationary point of the strictly concave function and hence is this maximum.
Only continuity of the price functions is given, so it is useful to justify trajectories without silently assuming differentiability. Put and . The dynamics become ordinary gradient flow ascent, . The function is strictly concave on the positive orthant: its user terms are plus constants, and each penalty is a convex nondecreasing function composed with the convex load . Hence the supporting-tangent inequalities give
The squared distance between two solutions is nonincreasing, proving uniqueness even though the vector field need only be continuous.
Nonnegative initial rates enter the positive orthant immediately: along a solution, variation of constants in expresses as a positive exponential times plus a strictly positive integral. Also , so rates cannot become infinite in finite time. The distance argument applies after any positive time and, by taking that time down to zero, also proves uniqueness for initial zero rates.
For a positive starting time, the increasing traps the trajectory in a compact superlevel set. Its derivative has finite integral since is bounded above there. It is uniformly continuous in time, being a continuous function of the state on a compact set with bounded velocity. By uniformly continuous integrable functions vanish at infinity, . Every limit point consequently has for every , since all rates are bounded away from zero. There is only the limit point , so
This establishes global convergence of primal congestion control with the needed nonzero-price condition, and supplies the zero-price counterexample when increasing is read in the weak sense.
For the alternative increasing-supply resource-price dynamics, consider the dual congestion potential
Its domain is with every . Each integrated supply is strictly convex, because is strictly increasing, and each negative logarithm of a positive linear function is convex. Thus is strictly convex. Its sublevel sets are bounded: for each , gives a linear lower bound on the supply primitive at large , whereas the negative logarithmic terms grow only logarithmically in . A route-price sum tending to zero makes on a bounded set. A minimum therefore exists in this domain.
The derivatives are
If a used resource had at the minimum, then but , so increasing that coordinate would strictly lower . Therefore the minimizer has strictly positive prices at every used resource. An unused resource has its unique minimizing price zero. At used resources its equations are
The multiplicative dynamics are . On strictly positive used-resource prices, equilibrium is equivalent to these equations, and strict convexity proves existence and uniqueness of the positive equilibrium. No full-row-rank assumption on is needed here: the strictly increasing individual supplies already give strict convexity.
If zero prices at used resources are allowed, the unqualified uniqueness assertion is false because of boundary equilibria of multiplicative resource prices. For one unit-weight route using two resources and , all three price vectors
are equilibria. At the user rate is one; the first resource has demand equal to supply, while the second has zero derivative because its price is zero. The last two vectors do not minimize . Thus the positive-price domain, or a boundary rule preventing a used resource from remaining at zero price, is essential.
Finally, the primal and dual equilibrium equations agree when and are inverse functions on the relevant ranges. For example, continuous strictly increasing prices with and have inverses satisfying the stated supply assumptions. Then is equivalent to , while both systems have . With inverse price and supply functions, the two equilibria describe the same rates and prices. Without this relation they need not agree: for one route and one resource with , and , the primal equilibrium is while the dual equilibrium is .
In congestion control, resource observes only its own aggregate load and generates the nonnegative feedback price . User receives the sum of prices along its route. Its desired increase competes with the congestion charge , while sets the adjustment speed. This is local information: a resource needs the rates passing through it, and a user needs the feedback along its own route, rather than the complete network state.
There is a hypothesis to make explicit before claiming global convergence. If “increasing” means nondecreasing, the printed assumptions allow , in which case and there is no equilibrium point. The intended theorem holds when every route contains at least one resource whose price is not identically zero, with nonempty routes and nonnegative, continuous, nondecreasing prices. Strictly increasing prices satisfy this condition automatically. We prove the theorem under this necessary qualification; without it the zero-price example disproves the literal claim.
Let . Each is a nonnegative convex function. For positive rates define the primal congestion potential
The weighted logarithmic utility is a strictly concave function, and subtracting the resource costs preserves strict concavity. Thus has at most one maximizer and at most one critical point on the positive orthant.
It also attains its maximum there. A nonzero nondecreasing price is bounded below by some positive constant at all sufficiently large loads; hence its primitive satisfies with . Choose at zero-price resources. The route coverage assumption gives , so
This bounds all superlevel sets of . Within a bounded set, approaching makes . Consequently every nonempty superlevel set is a compact set contained in the positive orthant. A maximizing sequence stays in one such set, giving a unique interior maximizer . Its critical-point equations are
The rate dynamics are the gradient flow with diagonal mobility . Along any positive trajectory,
with equality only at . Thus is a strict Lyapunov function. The initial superlevel set traps the trajectory away from both infinity and the boundary, ensuring global continuation. Positive initial rates remain positive: variation of constants in makes this explicit. Starting with finite nonnegative rates also gives positive rates at every later positive time.
For completeness, mere continuity of suffices for convergence; one need not silently assume differentiable prices. On the trapping compact set, both the vector field and are continuous and bounded. The trajectory satisfies a uniform Lipschitz condition in time, so is uniformly continuous. Also , since increases to a finite limit. The decay of a nonnegative uniformly continuous integrable function applies: otherwise disjoint intervals of a fixed positive width would each contribute a fixed positive amount. Every accumulation point therefore has and is . Compactness then proves the global convergence
This is the dissipation argument behind the LaSalle invariance principle. If uniqueness of trajectories is desired under continuous prices, put . The transformed equation is . The transformed potential is concave: it is a sum of positive multiples of , minus convex nondecreasing composed with nonnegative quadratic sums. Its gradient is monotone decreasing, so the squared distance between two solutions cannot increase. This proves uniqueness for positive initial data without a locally Lipschitz function hypothesis on .
For the alternative price dynamics, restrict first to positive prices on resources used by at least one route, and discard unused resources. Define the dual congestion potential
Its domain has and for every route. Strictly increasing make each strictly convex, and is convex. Thus is a strictly convex function. Each is eventually bounded below by a positive constant, so its integral grows at least linearly, dominating the negative logarithms as . If a route price tends to zero in a bounded set, . Therefore attains a unique minimum on its domain including the admissible boundary faces.
That minimum has every used-resource price positive. At a face , all route prices are still positive and the inward derivative is
Hence the minimum cannot lie on that face. At the unique interior minimum,
Since the price dynamics are , their positive equilibrium points are exactly these critical points. There is one equilibrium in the positive price domain. An unused resource instead has its only nonnegative equilibrium at zero.
The positive-domain qualification matters here even with strictly increasing . If zero prices are admitted, the multiplicative factor permits boundary equilibria of multiplicative price dynamics. Take one route using two resources, , and . Both and are equilibria, with , as is the positive equilibrium , with . Thus the literal assertion of uniqueness on the whole nonnegative orthant is false. A boundary equilibrium need not satisfy load equals supplied capacity at its zero-price resources.
The relationship between the two intended equilibria is inverse resource response. The first system has ; the positive equilibrium of the second has . If on the relevant ranges, these are the same resource equations, and both systems also impose . They therefore have the same rates and prices. For strictly increasing continuous with and unbounded range, the ordinary inverses satisfy all the stated conditions on . With arbitrary independently chosen and , the equilibria need not agree. Under inverse response, and are convex conjugates on the nonnegative half-line, and supplies the Lagrangian dual problem for maximizing the weighted logarithmic utility minus resource costs, up to constants independent of .
QUIC Created 2026-10-25 Updated 2026-11-05
QUIC is a transport protocol over User Datagram Protocol that provides multiple flow-controlled streams, cryptographic protection and connection migration. RFC 9000 published the version-1 transport specification in May 2021. Several streams can share one connection while retaining separate ordering: loss affecting one stream need not stop delivery of already available data on another. Congestion control and connection-wide limits still couple their use of network resources.[ref]
At the 17 January 2020 examination date, the cited November 2019 design was an Internet-Draft and work in progress. This historical description does not treat a later standard as already available.[ref]
For positive rates in congestion control, put and , where is the primal congestion potential. The equation becomes . Positive logarithmic utility weights make strictly concave on the positive orthant. Each integrated nonnegative nondecreasing resource price is a convex nondecreasing function of its load; composing it with the convex quadratic load in preserves convexity. Hence is a strictly concave function, even for prices that are merely continuous. For two solutions, . This proves uniqueness of the flow without assuming local Lipschitz continuity of the price functions.