FR2825546A1 - Packet switching system for constant matrix throughput feeds least filled destinations first - Google Patents

Abstract

A packet data switching system has a matrix (100) with source (200) and destination (400) queues with dispatch notification circuits (307) linked by switches (301) to multiplexers (300) with priority stored in memory (302) and read by queue selection circuits (309) to ensure that least filled destination queues are fed first.

Description

<Desc / Clms Page number 1>

 The present invention relates to a data packet switching system, particularly suitable for high speed telecommunications network equipment, making it possible to regulate the data flows in transit between its input terminals and its output terminals in order to guarantee quality. of service on said flows.

The control of the quality of service, commonly called QoS (Quality Of Service), is a major issue to offer the user multimedia services (voice, video, data, etc.) federated by a single telecommunications network. It can be described simply by transfer characteristics between the entry point and the exit point of a network or switching equipment: - A first quantity is linked to the transmission time and applies to time services -real. It breaks down into a fixed time mainly related to the transmission speed in the optical or electrical beams and a variable delay (jitter) caused by the instantaneous congestion of the packets requesting the same output terminals of the intermediate network equipment. It must be mastered in order to offer interactive services. Thus a total transmission time greater than a few hundred milliseconds makes any telephone conversation unpleasant or even impractical. Jitter is today relatively well controlled in switching matrices insofar as a few allowable jitter values are authorized, which in practice satisfies the vast majority of applications.
A few queues served as output according to fixed priorities allow satisfactory results to be obtained: The more the jitter allowed, the more the associated queue will have a high priority.

- A second magnitude is linked to the loss of data, which will result in a degradation of the service offered on a real-time application (alteration of the voice or video) and not real-time by the increase in retransmissions operated automatically by the upper layers of the protocols (increased transfer times on files). In order to control the effects, techniques for the controlled destruction of packets in queues are then used in conjunction with saturation prevention signals (called backpressure signals). The latter, usually issued by switching dies or cores, act in feedback on the queues placed at their entrances.

 A third quantity is linked to the speed guaranteed by the network to make a given service viable or comfortable. For example, any Internet user has experienced the inconvenience caused by the absence of a minimum service, resulting in

<Desc / Clms Page number 2>

   outright freeze on his Internet browsing. Controlling guaranteed rates is problematic, particularly when dies are used in switching equipment.

Readers can refer to the book Quality of Service in ATM Networks by Natalie Giroux & Sudhakar Ganti published in 1998 by Prentice Hall / USA to familiarize themselves with these notions of QoS and the main types of switching matrices.

The main object of the present invention is to overcome this problem of controlling guaranteed bit rates when switching matrices are used and relates to a switching system, suitable in particular for telecommunications network equipment, characterized essentially in that '' it includes: w a plurality of input terminals for receiving packet flows on which are guaranteed flow constraints, a plurality of output terminals for transmitting said flows on conduits specific to the telecommunications network, w a matrix switching capable of transmitting said packets from its input ports to its output ports as a function of routing information contained in the header of said packets, w execution blocks of the input functions of said network equipment, in particular responsible for recognizing said packets identifying said flows and providing the information necessary for ro use of said packets through said matrix, a plurality of source queues each receiving a stream of packets and placed upstream of said input ports of said matrix, w a plurality of destination queues each receiving a packet flow coming from one of said output ports of said matrix,. blocks for executing the output functions of said network equipment, in particular responsible for scheduling the data packets on each of the output terminals so that the frequency of extraction of the packets from said destination queues satisfies the constraint guaranteed flow of each flow,. a system for multiplexing packets from said source queues supplying said input ports of the matrix, characterized in that it comprises: w circuits for generating notifications signaling the output of packets from the queues of destination, said notifications being assigned routing information allowing their routing to the circuits of

<Desc / Clms Page number 3>

peut être sans mémoire de paquets interne de type crossbar/wormhole (ce terme peut se traduire en français par trou de vers car le tête du paquet commuté est arrivée dans la file de destination alors que la queue se trouve encore dans la file d'attente de source, évoquant ainsi l'image d'un vers), ou avec mémoire de paquets interne de type store and forward (avec une mémoire interne permettant de tamponner les paquets et disposée sur les entrées ou sur les sorties ou partagée entre les entrées/sorties). storage and information allowing the identification of their source queue with which they are associated in a one-to-one manner, storage circuits allowing the classification of said notifications, w selection circuits allowing the election of priority notifications and the selection of the source queues designated by said notifications,. a circuit for switching said notifications from said generation circuits to said storage circuits. The invention makes no assumptions about the internal structure of the packet switching matrix (commonly called switchingfabric). For example, this

can be without internal crossbar / wormhole packet memory (this term can be translated into French by wormhole because the head of the switched packet has arrived in the destination queue while the queue is still in the queue source, thus evoking the image of a worm), or with internal packet memory of the store and forward type (with an internal memory for buffering the packets and arranged on the inputs or on the outputs or shared between the inputs / exits).

The invention does not relate to the processing of real-time flows, the latter must therefore benefit, for example, from higher priority levels, from conventional use on matrices. In addition, the packets handled can be of fixed size in which case they are usually called cells as with the broadband ATM protocol (Asynchronous Transfer Mode) or of variable size in which case they are called frames as with the broadband IP protocols (Internet Protocol) and IEEE 802. x (Ethernet). The switching system which is the subject of the invention indifferently processes cells and data frames (like the most recent matrices).

- la figure 3 donne la méthode utilisée de régulation des flux sur un système de commutation conforme à l'invention, - la figure 4 donne le format du paquet transmis au travers du système de commutation, - la figure 5 présente un exemple de système de commutation conforme à l'invention. The principles used to solve the problem are presented and an embodiment of the invention is described below by way of example, with reference to the accompanying drawings in which: - Figure 1 shows the conventional use of a switching matrix, FIG. 2 shows the diagram of an ideal switching system without matrix,

- Figure 3 gives the method used to regulate flows on a switching system according to the invention, - Figure 4 gives the format of the packet transmitted through the switching system, - Figure 5 shows an example of a switching according to the invention.

<Desc / Clms Page number 4>

numérotés FI. K, l Q) par les blocs d'exécution des fonctions d'entrée 5001 K Ces derniers peuvent être en charge d'autres tâches comme le contrôle de trafic ou la gestion des files d'attente d'entrée de réception des paquets. Ces files dites de source 200] K, J Q mémorisent les paquets selon leur flux d'appartenance. Il est à noter que le nombre de files d'attente peut être de plusieurs centaines voire milliers. Les paquets sont ensuite introduits par les multiplexeurs 300 !... K dans la matrice 100 sur ces ports d'entrée 1101 K puis extraits des ports de sortie 120 !... L pour être mis dans des files d'attente de sortie dites de destination 400 !... L, 1. M. Les blocs d'exécution des fonctions de sortie 6001 L ordonnancent alors les paquets issus desdites files d'attente de destination vers les bornes de sortie 601 !... L du système de commutation. Les blocs 6001. L peuvent être en charge d'autres tâches comme le lissage de trafic par exemple. The conventional use of the switching matrix is shown in Figure 1: The packets arrive via the 5011 K input terminals of the system to be identified (by flow

numbered FI. K, l Q) by the execution blocks of the input functions 5001 K These can be responsible for other tasks such as traffic control or management of packet reception input queues. These so-called source queues 200] K, JQ store the packets according to their membership flow. It should be noted that the number of queues can be several hundreds or even thousands. The packets are then introduced by the multiplexers 300! ... K into the matrix 100 on these input ports 1101 K then extracted from the output ports 120! ... L to be put in so-called output queues destination 400! ... L, 1. M. The execution blocks of the output functions 6001 L then schedule the packets coming from the said destination queues towards the output terminals 601! ... L of the switching. Blocks 6001. L can be in charge of other tasks like smoothing traffic for example.

According to the example of FIG. 1, the problem posed by the matrix appears on one of the output ports at 11: the flow convergence 10L. w a total flow higher than that of the outlet port causes congestion. Even in the presence of conventional feedback mechanisms to limit the overall introduction of packets to the matrix input ports for a given output port, the packet scheduling device disposed on each of the matrix output ports is too spartan and remains unable to balance this introduction according to the guaranteed flow of each flow because without precise knowledge of their individual constraints. It generally implements a few fixed priorities with a rotating priority on the same priority level. This therefore results in a degradation of the quality of service even if the 6001 L blocks have the qualities necessary to ensure a guaranteed throughput on each flow.

Figure 2 shows the diagram of an ideal congestion-free switching system with a single output port s which will help the reader to understand the founding and underlying idea of the invention: Here we focus on a given flow carried by any queue 700u, v with u element from the set [1 ... K] of input ports and v element from the set [1 ... Q] of the queues. It is observed that the quantity of packets present in this queue is not necessary and useful information to guarantee a throughput on the associated stream. A simple boolean value of absence / presence of packets in the queue is sufficient for the block 600s to fulfill its quality of service objectives, which is notified of the introduction of packets in the queues by the link 602s . The 700u, v queue can be broken down into a couple of queues

<Desc / Clms Page number 5>

de destination : il en découle que si la file d'attente de destination 4005, t contient toujours au moins un paquet alors la valeur booléenne ( d'absence/présence de paquets dans la file d'attente 4005, t) nécessaire et suffisante au bloc 6005 est maintenue idéalement, on peut donc considérer que l'introduction d'une matrice n'a pas d'incidence sur la qualité de service. source / destination (200u, v 4005, t) according to the representation of figure 1 with s element of the set [1 ... L] of the output ports and t element of the set [1 ... M] queues

destination: it follows that if the destination queue 4005, t always contains at least one packet then the Boolean value (absence / presence of packets in the queue 4005, t) necessary and sufficient for the block 6005 is ideally maintained, so we can consider that the introduction of a matrix does not affect the quality of service.

1) Toute variation négative-c'est à dire toute sortie de paquet d'une file d'attente de destination 4005, t-est compensée au plus tôt par l'extraction et l'émission d'un paquet de la file d'attente de source 200u, v de la part du multiplexeur 300u. Pour cela un message, appelé notification, est renvoyé au multiplexeur 300u pour provoquer cette opération. Cette action s'oppose à la cause qui la produit, de sorte que les débits garantis sur les flux par les blocs 6001. L en aval de la matrice, se trouvent rétro- propagés à son amont. The main objective of the switching system according to the invention is to ensure that the destination queues 4001. L, t .. M are always supplied and therefore in a non-empty state (unless of course the associated source files are empty), for this:

1) Any negative variation - i.e. any packet output from a destination queue 4005, t - is compensated as soon as possible by the extraction and transmission of a packet from the queue 200u, v source waiting from multiplexer 300u. For this a message, called notification, is returned to the multiplexer 300u to cause this operation. This action is opposed to the cause which produces it, so that the guaranteed flows on the flows by the blocks 6001. L downstream of the matrix, are back propagated upstream.

2) The multiplexer 300u can have, at a given instant, a multitude of notifications to be processed and therefore destination queues to be filled. It therefore ranks on notifications so that the least crowded destination queues are replenished first. For this, the notifications carry priority information relating to the filling level of the destination queue that they represent.

3) The destination queue acts as a cache memory and for high speed or very sporadic flows, it may be necessary that a fairly large number of packets be provisioned in order to compensate for the replenishment delay introduced by the multiplexer 300u and the matrix 100. A high priority notification, signaling an unoccupied destination queue (typically empty) is then interpreted by the multiplexer 300u by the transmission of several packets (typically 2) for the destination queue considered. The number of packets, necessary and provisioned in a destination queue so that it is never empty, increases by adjusting itself automatically and quickly in a dedicated manner according to each flow. Conversely, a notification of

<Desc / Clms Page number 6>

 low priority, representative of a busy destination queue, results in the transmission of a single packet, satisfactory action according to 1) FIG. 3 graphically gives the regulation method which has just been introduced. The packets of a given stream arrive in the source queue 200u, v (identified by v and connected to port u of the matrix) by the link 502u. They are only authorized to exit when the scheduler 309u activates the extraction and the emission of these thanks to the command 303u which activates the contactor 20lu, v shown here symbolically.

commutateur de notifications 301 et à sa réception, la notification 23 formée du triplet (u, v, Pi) est classée dans 302u et déposée dans une file d'attente de notifications P, notée 308u, i (Pi est un élément de l'ensemble PO. N des files d'attente de notifications) en fonction de sa priorité Pi de la notification. A chaque file d'attente de notifications correspond un niveau de priorité de notifications. When a packet is extracted from the destination queue 400s, t (identified by t and connected to the port s of the matrix), a notification 23 is returned to the set of notification queues 302u via of a small notification switch 301. The switching capacity of the latter is smaller than that of the matrix 100 due to the size of the notifications (of the order of 32 binary elements) much smaller than that of the data packets ( from 400 to 10,000 binary elements). However, this capacity must nevertheless be sufficient to overcome congestion problems. Any notification carries a priority level Po. N dynamically evaluated by the block 307s as a function of a series of filling thresholds placed on the destination queue 400s, t. For example, priority P will be chosen when between 2 ′ and 2 ′ + ′ - l packets are present in said queue at the time preceding the extraction. After its routing by the

notification switch 301 and on receipt, the notification 23 formed of the triplet (u, v, Pi) is classified in 302u and deposited in a queue of notifications P, denoted 308u, i (Pi is an element of the set PO. N of the notification queues) according to its priority Pi of the notification. Each notification queue has a notification priority level.

At output, the scheduler 309u chooses the notification (u, v ', Pj) at the head of the highest priority queue not empty in the set of notification queues 302u (Po: highest priority, PN : lower priority) and sends the packet at the head of the source queue identified by v 'as soon as the input port of the matrix is available. In the event that no packet is no longer present in the source queue, the notification is destroyed; if in addition, the priority of the notification denotes the total absence of packets in the destination queue (revealed by Po) then the flow is said to be defused and boolean information Au, v, specific to the flow considered, is inactivated in the Priming table 304u. If the priority of the notification is maximum while the source queue is not empty, then the scheduler 309u extracts several packets

<Desc / Clms Page number 7>

 (typically 2) at the head of it and emits them, to the extent of course of their availability.

When a packet is introduced into the source queue 200u, v when the latter is empty and Au, v in the inactive state, Au, v is activated by the event 21 and a notification (u, v, Po) is sent to the queue set 302u on link 22: The flow is then said to have started.

Figure 4 gives an example of the packet format transmitted within the switching system. The header 30 contains the fields necessary for switching the packet: The information provides information on the output port of the matrix, t represents the identifier of the destination queue. In the same way, u provides information on the input port of the matrix and v represents the identifier of the source queue. The packet Pj field represents the priority level of the notification which triggered the transmission of said packet.

Finally, field 31 contains the useful data switched between the input and output terminals of the system.

The simulations carried out show that the switching system which is the subject of the invention works correctly without any particular arrangement of the matrix 100. According to a characteristic of the invention, said system can nevertheless take advantage of the possibilities thereof to improve its efficiency. In this sense and if the matrix is of the type without internal memory, the latter orders the packets in congestion according to the priority carried by the field Pj of the packets on each of its output ports. This measure is not applicable with matrices having an internal memory capable of storing a multitude of packets. Indeed, the priority of the packets of a flow is variable and changes according to the filling level of the destination queue of said packets, the latter therefore risk arriving disordered in said queue, certain packets doubling others in the matrix .

According to another characteristic of the invention applicable to a matrix having an internal memory, a graduated level of internal saturation of said matrix available on its input ports modifies the behavior of the multiplexer 300u so that it ignores the queues waiting for low priority notifications when the matrix is congested. Thus, the supply of lightly filled destination queues is de facto a priority in the matrix. The election of a notification is made by 309u in the queues of notifications of priority higher than a threshold fixed dynamically according to the level of saturation of said matrix according to the rule

<Desc / Clms Page number 8>

 The higher the saturation, the more the election is restricted to queues of high priority notifications.

FIG. 5 gives a schematic view of an example of a switching system according to the invention. The data packet switch 100 and the notification switch 301 are combined in a single matrix. The notifications are transported individually in packets of a homogeneous format with the payload data packets so that the matrix switches between the two types of information. It is of course desirable for this matrix to process packets of variable size so that the load caused by notification transfers remains negligible compared to that of useful data packets.

Each pair of input / output ports of the matrix is connected to a coupler card 900 detailed here according to the index u supporting the functions necessary for its connection to conduits 501u / 601u specific to the telecommunications network. The 500u / 600u network function blocks could of course individually manage a plurality of said conduits.

The 111u multiplexer with two inputs makes it possible to send the useful data packets present at the output of the set of source queues 200u, 1 Q present on 112u and the notifications produced by 307u and presented on 310u to the same port. llOr matrix input. It is desirable that notifications have priority over useful data packets.

From the same output port 120u in the matrix, the same information is demultiplexed by the element 121u to respectively the block 300u for notifications via 320u, and the set of destination queues 400u, l ... M for useful packages via 122u. A boolean value in the header of the packets makes it possible to distinguish whether they contain useful data or notifications. Of course, separate pairs of input / output ports for notifications and useful data packets are also possible.

The set of destination queues 400u, l M is integrated into a so-called destination memory where dynamic sharing is carried out on the space allocated to each queue because, considering M queues which can contain at most H packets, it it is very unlikely that M * H places will be needed to store the packets in said memory. The maximum size of a queue can be evaluated dynamically based on the memory usage rate. Considering F the number of places

<Desc / Clms Page number 9>

restant disponibles dans la mémoire pour recevoir des paquets, la taille maximale de la file d'attente H peut, par exemple, s'exprimer ainsi : H = a * (F-S) où a est un coefficient d'allocation compris entre 0 et 1 (typiquement en l/2n), et S une marge de sécurité liée au fait que des notifications associées à la mémoire peuvent déjà être mémorisées par le multiplexeur 300. Ainsi, plus la place vient à manquer dans la mémoire, plus la taille maximale des files d'attente se réduit ; le partage de la mémoire reste équitable empêchant ainsi que certaines files d'attente ne monopolisent trop la mémoire. Si la taille courante T d'une file d'attente vient à dépasser la taille maximale H, alors aucune notification n'est générée par le bloc 307 ; ce mécanisme permet ainsi de se préserver d'une saturation de la mémoire de destination qui entraînerait une perte de paquets et par conséquent serait préjudiciable à la qualité de service.

remaining available in the memory to receive packets, the maximum size of the queue H can, for example, be expressed as follows: H = a * (FS) where a is an allocation coefficient between 0 and 1 (typically in l / 2n), and S a safety margin linked to the fact that notifications associated with the memory can already be memorized by the multiplexer 300. Thus, the more space runs out in memory, the more the maximum size of the queues get reduced; memory sharing remains fair, preventing certain queues from monopolizing memory too much. If the current size T of a queue exceeds the maximum size H, then no notification is generated by block 307; this mechanism thus makes it possible to protect itself from saturation of the destination memory which would cause packet loss and therefore would be detrimental to the quality of service.

L'introduction d'une trame de longueur Le dans la file d'attente de destination y se traduit sur Tj par : Tj = Tj + Le. L'extraction d'une trame de longueur Ls dans le file d'attente de destination y se traduit sur Tj par : Tj = Tj-Ls. La priorité de notification est toujours évaluée en fonction de T : par exemple la priorité Pi est choisie lorsque la file d'attente contient entre T=2'*, 6et T= (21+1-1) *, 6 éléments binaires où ss représente un coefficient de quantification (typiquement en 2n). Ces calculs sont homogènes avec des paquets de taille fixe (où Le=Ls= 1) et se présentent alors comme un simple comptage de paquets. When the payload data packets are of variable size, the filling level of a destination queue is not expressed only in terms of number of packets but is weighted by the length of said packets. Said size T is then expressed by the cumulative length (for example expressed in number of binary elements) of the packets present in the destination queue considered. Since the output terminal has a flow D (in binary elements per second), the provision made in the destination queue is expressed by T / D and is homogeneous at one time.

The introduction of a frame of length Le in the destination queue there results on Tj by: Tj = Tj + Le. The extraction of a frame of length Ls in the destination queue y results on Tj by: Tj = Tj-Ls. The notification priority is always evaluated as a function of T: for example the priority Pi is chosen when the queue contains between T = 2 '*, 6 and T = (21 + 1-1) *, 6 binary elements where ss represents a quantification coefficient (typically in 2n). These calculations are homogeneous with packets of fixed size (where Le = Ls = 1) and are then presented as a simple counting of packets.

According to a variant of the invention where an optimization device - with which the invention should not be confused - is implemented in the matrix (the crossbar type has a maximum load limited to 58% without optimization), each port the input port of the latter usually has a virtual queue upstream for each of the output ports, that is to say by direction (technique commonly called VOQ: Virtual Output Queuing). The optimization device governs the connections established between input / output ports at all times and can thus avoid packets on N

<Desc / Clms Page number 10>

ports d'entrée différents ne sollicitent simultanément la même sortie conduisant à une situation d'attente de N-1 paquets et de blocage sur N-1 entrées. Le lecteur pourra se référer au document < Scheduling Algorithms for Input queued cells switches > ) McKeown, Ph. D. thesis, University of California, Berkeley, May 1995 pour se familiariser avec ces méthodes. Le dispositif d'optimisation est alors utilisé conjointement avec le bloc 300 et fournit à ce dernier le port de sortie de matrice à servir et sur lequel une connexion est réalisée. Le bloc 300 dispose, par port d'entrée u, d'une structure 302u, s de files d'attente de notifications par port de sortie s (soit au total L blocs 302u par entrée u) de sorte que l'élection d'une notification est faite par 309u dans la structure 302u, s relative au port de sortie fourni par le dispositif d'optimisation. Des moyens semblables sont à mettre en oeuvre lorsque la matrice, de type store and forward, fournit des signaux de saturation relatifs à chacun de ses ports de sortie. Dans ce cas, l'élection d'une notification est faite par 309u dans la liste des structures 302u, s (relatives aux ports de sortie non saturés) fournie par le dispositif d'optimisation.

Different input ports do not simultaneously request the same output, leading to a waiting situation for N-1 packets and blocking on N-1 inputs. The reader can refer to the document <Scheduling Algorithms for Input queued cells switches>) McKeown, Ph. D. thesis, University of California, Berkeley, May 1995 to familiarize themselves with these methods. The optimization device is then used in conjunction with the block 300 and provides the latter with the matrix output port to be served and to which a connection is made. Block 300 has, by input port u, a structure 302u, s of notification queues by output port s (ie in total L blocks 302u per input u) so that the election of a notification is made by 309u in the structure 302u, s relating to the output port provided by the optimization device. Similar means are to be used when the matrix, of the store and forward type, provides saturation signals relating to each of its output ports. In this case, the election of a notification is made by 309u in the list of structures 302u, s (relating to the unsaturated output ports) provided by the optimization device.

Network processors, traditionally micro-programmed to execute the functions of blocks 500 and 600, can also support the multiplexing system 300 and 307 specific to the switching system object of the invention. Said multiplexing system can also be integrated into FPGAs (Field Programmable Gates Array) or to specific circuits of the ASIC (Application Specific Integrated Circuit) type.

Claims (11)

1. Data packet switching system, particularly suitable for high speed telecommunications network equipment, making it possible to regulate the data flows in transit between its input terminals and its output terminals in order to guarantee a flow rate on said flows , and which is essentially characterized in that it comprises:. a plurality of input terminals (501) for receiving packet flows on which are imposed guaranteed bit rate constraints, w a plurality of output terminals (601) for transmitting said flows on conduits specific to the telecommunications network, a switching matrix (100) capable of transmitting said packets from its input ports to its output ports (120) according to routing information contained in the header of said packets, execution blocks (500) of the input functions of said network equipment, in particular responsible for recognizing said packets identifying said flows and providing the information necessary for routing said packets through said matrix, w a plurality of source queues (200 ) each receiving a packet stream and placed upstream of said input ports of said matrix,. a plurality of destination queues (400) each receiving a stream of packets originating from one of said output ports of said matrix, execution blocks (600) of the output functions of said network equipment, in particular responsible for the scheduling of the data packets on each of the output terminals so that the frequency of extraction of the packets from said destination queues satisfies the guaranteed rate constraint of each of the flows, w a system for multiplexing the packets originating from said source queues supplying said input ports of the matrix, characterized in that it comprises: w circuits for generating (307) notifications signaling the output of packets from destination queues, said notifications being assigned routing information allowing their routing to the storage circuits and information allowing the identification of their source queue with which it are associated in a unique way, <Desc / Clms Page number 12>   storage circuits (302) allowing the classification of said notifications in notification queues, selection circuits (309) allowing the election of priority notifications and the selection of source queues designated by said notifications, w a switching circuit (301) of said notifications from said generation circuits to said storage circuits, 2. Switching system according to claim 1, characterized in that a notification is generated following the extraction of a packet from a destination queue, 3. Switching system according to claims 1 and 2, characterized in that the generation of a notification is inhibited when the maximum filling level of the associated destination queue is reached, 4. Switching system according to claims 1 to 3, characterized in that the notifications also carry priority information linked to the filling level of their associated destination queue, a low filling level conferring a high priority, 5. Switching system according to claims 1 to 4, characterized in that the notifications are stored in queues according to the priority level of said notifications received by the storage circuits, 6. Switching system according to claims 1 to 5, characterized in that a selection circuit chooses, to transmit on the matrix input port to which it is connected, the packet at the head of the source queue whose the associated notification is at the head of the highest priority non-empty notification queue, 7. Switching system according to claims 1 to 6, characterized in that a selection circuit transmits several packets when the priority of the notification, associated with the source queue from which they originate, is greater than a predefined threshold. , 8. Switching system according to claims 1 to 7, characterized in that a selection circuit destroys a notification associated with a source queue that has become empty, 9. Switching system according to claims 1 to 8 and a first variant, characterized in that: <Desc / Clms Page number 13>  'the storage circuits rank the notifications by priority level and additionally according to the matrix output ports used by the packets of the source queues with which they are associated,' the selection circuits operate the election of notifications by priority level on a list of matrix output ports, said list being dynamically provided and dedicated by an optimization device to each of said circuits, 10. Switching system according to claims 1 to 9 and a second variant applicable to a matrix without packet memory, characterized in that the packets carry the priority level of the notification which has led to their extraction from the source queues so that the packets associated with the least filled destination queues are treated as a priority within said matrix, 11. Switching system according to claims 1 to 9 and a third variant applicable to a matrix having a packet memory, characterized in that the selection circuits take into account in their election phase only the queues waiting for notifications of high priority when memory saturation information is emitted by said matrix.