DE10148893A1 - Method for transmitting streams of traffic in a connectionless packet-oriented communications network uses transmission nodes interconnected so as to create multiple paths between the nodes. - Google Patents

Abstract

A connectionless packet-oriented communications network (300) has multiple transmission nodes (301-315) interconnected so as to create multiple paths between the transmission nodes. One part of the nodes forms input nodes (301-307,E) and/or output nodes (301-307,F) for a communications network. The input nodes send streams of traffic to the output nodes. More than one residual path to an output node is detected for one of the transmission nodes receiving part of the transmitted streams of traffic. This transmission node performs distributed transmission onto two of the detected residual paths for the streams of traffic sent to the output node. Independent claims are also included for a device for carrying out the method of the present invention and for a system like a communications network for this device.

Description

In a communication network, the traffic should follow certain rules as evenly as possible on all nodes and Connection lines in the communication network - also called "network" - be distributed. The network cannot from that in the Pre-published German patent application DE 101 49 349.9 be disclosed type, the disclosure of which by reference to the content the present description is made.

This type of distribution results in a so-called “distribution fan” for each communication relationship from a specific input A to a specific output B, which comprises all nodes and connection paths that can be usefully used for this communication relationship (see FIG. 1). In a correspondingly meshed network, the distribution subjects of different communication relationships inevitably overlap in such a way that either identical or partially overlapping or also completely disjoint "branching patterns" arise at the individual nodes (see FIG. 2). The overlap depends on the distribution mechanisms used in the network nodes.

• 1. Einfache Verteilung des ankommenden Verkehrs auf ein abgehendes Bündel ohne Prioritäten:
  • a) Vorabverteilung des Verkehrs in individuelle Warteschlangen pro Port:
    Ein zentraler Verkehrsverteiler verteilt den ankommenden Verkehr auf einzelne Warteschlangen, von denen jeweils eine genau einem abgehenden Port des Bündels zugeordnet ist. Die Verteilung kann zyklisch oder (z. B. bei unterschiedlichen Port- Bandbreiten) gewichtet nach verschiedenen Kriterien erfolgen. Dabei kann z. B. auch der aktuelle Füllstand der einzelnen Warteschlangen (nach der Zahl der Pakete oder bei variablen Paketlängen nach der echten Datenmenge in Bytes) oder die individuelle Länge des aktuell zuzuteilenden Datenpaketes berücksichtigt werden. Die Ports bedienen die Warteschlangen in der Regel nach dem FIFO-Prinzip. Was einmal in einer Queue steht, muss dann auch vom zugeordneten Port abgearbeitet werden.
  • b) Verwendung einer einzigen Warteschlange mit einem Multi Server Prinzip:
    Eine günstige Verteilung des Verkehrs mit gleichzeitig optimaler Ausnutzung der verfügbaren Portkapazitäten kann mit dem Multi-Server-Prinzip erreicht werden. Dabei werden alle ankommenden Datenpakete in eine einzige Warteschlange eingereiht, aus der sich die Ports, wann immer sie frei sind bzw. frei werden, meist nach einem FIFO (First In First Out) Prinzip das nächste zu bedienende Paket abholen.
• 2. Verteilung des ankommenden Verkehrs auf ein abgehendes Bündel mit Prioritäten
  • a) Vorabverteilung des Verkehrs in individuelle Prioritäts-Warteschlangen pro Port:
    Ein zentraler Verkehrsverteiler verteilt den ankommenden Verkehr auf einzelne Warteschlangen, wobei für jeden abgehenden Port pro Prioritätsklasse jeweils eine individuelle Warteschlange bereitgestellt wird. Die Varianten gemäß 1 (a) sind analog anwendbar. Bei der Bedienung der Warteschlangen berücksichtigen die Ports die Prioritäten nach entsprechenden Regeln ("strikt", "gewichtet" usw.)
  • b) Multi Server Prinzip mit je einer Warteschlange pro Prioritätsklasse:
    Wie 1 (b), wobei die höherprioren Warteschlangen entsprechend der Prioritätsregeln bevorzugt bedient werden.
• 3. Verteilung des ankommenden Verkehrs auf ein abgehendes Bündel mit prioritätsgesteuertem Per Flow Queueing:
  Eine Verfeinerung der oben beschriebenen elementaren Mechanismen besteht darin, zur granulareren Unterscheidung und Priorisierung zwischen verschiedenen individuellen Kommunikationsbeziehungen (Flows) auch individuelle und separate Queues per Flow einzurichten. Damit wird jedoch der Aufwand für Queueing (wg. Zahl der Warteschlangen) und Scheduling (wg. Auswahl des nächsten auszugebenden Paketes aus der Vielzahl der Warteschlangen) um ein Vielfaches erhöht und sehr viel mehr von den Verkehrsmustern (d. h. Zahl der gleichzeitig aktiven Flows) abhängig. Zusätzlich muss dabei auch auf eine faire Verteilung der Ressourcen zwischen gleich priorisierten Flows geachtet werden, wofür spezielle Mechanismen wie z. B. "Weighted Fair Queueing" (WFQ) oder ähnliches eingesetzt werden, deren Komplexität (vor allem bei einer sehr großen Zahl von Queues) die eines einfachen Priority-Queueings um ein Vielfaches übersteigen kann. Selbstverständlich lässt sich Per Flow Queueing sowohl portindividuell [obiges Muster (a)] als auch in Verbindung mit dem Multi-Server-Prinzip [obiges Muster (b)] einsetzen.

So far, the following mechanisms for the individual distribution of data packets to outgoing bundles are known:

• 1. Easy distribution of incoming traffic to an outgoing bundle without priorities:
  • a) Advance distribution of traffic in individual queues per port:
    A central traffic distributor distributes the incoming traffic to individual queues, one of which is assigned to exactly one outgoing port of the bundle. The distribution can be cyclical or (e.g. with different port bandwidths) weighted according to various criteria. Here, for. For example, the current fill level of the individual queues (based on the number of packets or, in the case of variable packet lengths, the actual amount of data in bytes) or the individual length of the data packet currently to be allocated can be taken into account. The ports usually serve the queues according to the FIFO principle. What is in a queue must also be processed from the assigned port.
  • b) Using a single queue with a multi server principle:
    A favorable distribution of traffic with optimal utilization of available port capacities can be achieved with the multi-server principle. All incoming data packets are placed in a single queue, from which the ports, whenever they are free or become free, usually collect the next packet to be operated according to a FIFO (First In First Out) principle.
• 2. Distribution of incoming traffic to an outgoing bundle with priorities
  • a) Advance distribution of traffic into individual priority queues per port:
    A central traffic distributor distributes the incoming traffic to individual queues, with an individual queue being provided for each outgoing port for each priority class. The variants according to 1 (a) can be used analogously. When operating the queues, the ports take the priorities into account according to the corresponding rules ("strict", "weighted" etc.)
  • b) Multi server principle with one queue per priority class:
    Like 1 (b), but the higher priority queues are served preferentially according to the priority rules.
• 3. Distribution of incoming traffic to an outgoing bundle with priority-controlled per flow queuing:
  A refinement of the elementary mechanisms described above consists in setting up individual and separate queues per flow for more granular differentiation and prioritization between different individual communication relationships (flows). However, this increases the effort for queuing (due to the number of queues) and scheduling (due to the selection of the next packet to be output from the large number of queues) and is much more dependent on the traffic patterns (ie number of simultaneously active flows) , In addition, attention must also be paid to a fair distribution of resources between flows with the same priority. B. "Weighted Fair Queuing" (WFQ) or the like are used, the complexity (especially with a very large number of queues) can exceed that of a simple priority queuing by a multiple. Of course, per flow queuing can be used individually for each port [sample (a) above] as well as in connection with the multi-server principle [sample (b) above].

With all of the methods mentioned, the implementation in queues are usually just a pointer (Address) to identify the respective data packet in one usually shared data storage. The The order of operation results implicitly from the order the entries in the queue (e.g. after the FIFO Principle) or from the upstream process for selecting the as next queue to be served (e.g. by priority and with the same priority z. B. cyclical, longest queue first, shortest queue first, after weighting as with WFQ).

To achieve additional special effects, other Information included in this scheduling decision become. Especially in ATM technology is very common (but occasionally also in the IP environment) of the need for one Traffic shaping spoken. With this procedure in usually certain bandwidths, mostly realized by corresponding distance criteria between the cells (packages) a connection (i.e. communication relationship) become. This will include additional time information saved, the earliest, latest and / or optimal Service time for a queue or a specific one Specify cell (package) (so-called "calendar").

The mechanisms can also be used with the same effectiveness extend the application to several bundles, if these Bundles either do not overlap (are disjoint) or completely overlap (are identical). On Approach for a fair and efficient package-wise Traffic distribution on partially overlapping bundles with Priorities are not known.

An object of the invention is to show how in the network node taking into account a possible intended prioritization of individual traffic flows or individual data packets traffic according to given Branch patterns as optimal as possible on the outgoing Connection lines can be distributed, each network node decide autonomously and individually for each data packet.

This object is achieved by the invention. The invention provides that my network node per current branching pattern at least one queue each is provided. A queue may well be for multiple distribution compartments that are located in this node map the same branching pattern, be shared. Incoming packets become one according to their affiliation certain distribution subjects (i.e. not after the Belonging to a specific flow within the Distribution fan) distinguished. Accordingly, they are in one associated queue entered.

The queues change from one to the other Served ports belonging to the branching pattern. As soon as a port is free he chooses the one to be operated next Queue out. The selection starts on time there is no gap in the outgoing line. Serves a port several (in this case, in whole or in part overlapping) branching pattern, so this selection is first after not clearing all of these branching patterns Queues wanted. If several are not empty Queues found from different branching patterns, so a decision is made based on a predetermined criterion Queue is served next. That from this Queue to be issued packet is then added to another Criterion determined.

The number of queues required is advantageous here limited and only by the topology of the network, i. H. the Number of neighboring nodes considering one Bundling of ports, but not dependent on traffic.

There are no problems with a mixed use Networks since the invention is only local in a given Network node is used.

• - Zur Delay Optimierung werden mehrere, mit unterschiedlicher Priorität gekennzeichnete Warteschlangen innerhalb eines Verzweigungsmusters vorgesehen. Die Warteschlangen innerhalb eines Verzweigungsmusters werden strikt nach deren Priorität abgearbeitet. Bedient ein Port mehrere (in diesem Falle sich ganz oder teilweise überlappende) Verzweigungsmuster, so wird zunächst über alle diese Verzweigungsmuster nach der höchstprioren, nicht leeren Warteschlange gesucht und diese wird bedient. Werden dabei mehrere nicht leere Warteschlangen derselben Priorität aus verschiedenen Verzweigungsmustern gefunden, so erfolgt die Arbitrierung zwischen diesen Warteschlangen nach einem frei wählbaren Kriterien, z. B. dem Zufallsprinzip, einer zyklischen Reihenfolge, der Größe der Bündel, der Warteschlangenlänge (longest/shortest first), der vergangenen Zeit seit der letzten Bedienung dieser Warteschlange, der Anzahl der Ports, die für ihre Bedienung zuständig sind. Durch den Einsatz priorisierter Warteschlangen wird mit vergleichsweise geringem Aufwand eine ausreichend gute Delay- und Verteilungsqualität bewirkt.
• - Zur Optimierung auf den minimal möglichen Delay für jedes Paket werden ebenfalls mehrere, mit unterschiedlicher Priorität gekennzeichnete Warteschlangen innerhalb eines Verzweigungsmusters vorgesehen, die innerhalb eines Verzweigungsmusters nach strikter Priorität abgearbeitet. Bedient ein Port mehrere (in diesem Falle sich ganz oder teilweise überlappende) Verzweigungsmuster, so wird zunächst über alle diese Verzweigungsmuster nach der höchstprioren nicht leeren Warteschlange gesucht und diese wird bedient. Werden dabei mehrere nicht leere Warteschlangen derselben Priorität aus verschiedenen Verzweigungsmustern gefunden, so wird zwischen diesen nach einem Zeitkriterium entschieden. Dazu wird in den Warteschlangen neben einem Speicherzeiger (Adresse, s. o.) eine zusätzliche, vorzugsweise relative, Zeitinformation ("TimeStamp") mit abgelegt, aus der abgelesen werden kann, wann das Paket in die Warteschlange eingetragen wurde bzw. wie lange es sich schon darin befindet (Bild 4). Ausgewählt wird dann die Warteschlange, an deren Spitze sich das schon am längsten wartende Paket befindet, wodurch als Kriterium ein FIFO Prinzip realisiert wird.
  Mit dieser Ausgestaltung wird vorteilhaft sichergestellt, dass mit jedem frei werdenden Port, auch bei sich nur teilweise überlappenden Bündeln, über alle von diesem Port bediente Verzweigungsmuster hinweg immer die momentan gerade dringendste Anforderung mit der minimal möglichen Verzögerung bedient wird. Dies ist besonders dann sehr vorteilhaft, wenn als hochpriore Verkehrströme Daten interaktiver Echtzeitdienste, z. B. Telephon- oder Videokonferenzdienste, übermittelt werden.
• - Das paketindividuelle Zeitkriterium wird bei allen Warteschlangen verwendet, sondern nur bei hochprioren Verkehr (bspw. zeitkritischen Echtzeitanwendungen), um diesen mit dem kleinstmöglichen Delay und der entsprechend fairen Verteilung abzufertigen. Warteschlangen, die Verkehr ohne spezielle Delayanforderungen oder einfach nur Best Effort- Verkehr bedienen, können niedriger priorisiert und ohne Berücksichtigung dieser Zeitinformation behandelt werden. Prinzipiell reicht es aus, die Zeitinformation nur für die entsprechend hochprioren Warteschlangen zu speichern und auszuwerten.

Further developments of the invention are as follows:

• - For delay optimization, several queues with different priority are provided within a branch pattern. The queues within a branch pattern are processed strictly according to their priority. If a port serves several (in this case completely or partially overlapping) branching patterns, all of these branching patterns are first searched for and the queue is searched for the highest priority, not empty queue. If several non-empty queues of the same priority from different branch patterns are found, the arbitration between these queues is based on a freely selectable criteria, e.g. B. the random principle, a cyclical order, the size of the bundle, the queue length (longest / shortest first), the elapsed time since this queue was last operated, the number of ports that are responsible for its operation. By using prioritized queues, a sufficiently good delay and distribution quality is achieved with comparatively little effort.
• - To optimize for the minimum possible delay for each packet, several queues with different priority are also provided within a branch pattern, which are processed within a branch pattern according to strict priority. If a port serves several (in this case completely or partially overlapping) branching patterns, all these branching patterns are first searched for and the queue is used, which is not the highest priority. If several non-empty queues of the same priority are found from different branch patterns, a decision is made between them based on a time criterion. For this purpose, in addition to a memory pointer (address, see above), additional, preferably relative, time information (“TimeStamp”) is also stored in the queues, from which it can be read when the packet was entered in the queue or how long it had been there located ( picture 4). The queue is then selected, at the top of which is the longest waiting packet, whereby a FIFO principle is implemented as a criterion.
  This configuration advantageously ensures that with each port that becomes free, even with bundles that only partially overlap, the currently most urgent request is always served with the minimum possible delay across all branching patterns served by this port. This is particularly advantageous if data of interactive real-time services, e.g. B. telephone or video conferencing services are transmitted.
• - The packet-specific time criterion is used for all queues, but only for high-priority traffic (e.g. time-critical real-time applications) in order to process this with the smallest possible delay and the appropriate fair distribution. Queues that serve traffic without special delay requirements or simply best effort traffic can be given a lower priority and without taking this time information into account. In principle, it is sufficient to save and evaluate the time information only for the corresponding high-priority queues.

The granularity of time information should be the minimum Transmission time of a packet, i. H. the shortest possible Package on the fastest line of the bundle, consider. The possible range of values should be chosen in this way be that an overflow within the expected maximum delay time of a correspondingly high priority Package does not occur (taking into account the maximum lengths the associated queues, port bandwidths etc.) and a safe decision is possible at any time. It can also a possibly set limitation of the high priority Traffic in the network are taken into account with the advantage shorter queues.

• 1. Das Zeitkriterium wird auch dazu verwendet, um Pakete, die aus irgendwelchen Gründen zu lange gewartet haben und deren weitere Übertragung (im Rahmen der zugehörigen Echtzeitanwendung) nicht mehr sinnvoll erscheint, auszusortieren und zu verwerfen. Dies kann entweder beim Schedulen oder durch einen separaten, parallel laufenden Prozess erfolgen. Ein solcher Prozess könnte z. B. die Zeitinformation beim Eintragen des Paketes in die Warteschlange auf einen definierten Anfangswert zu setzen, sie regelmäßig getaktet auf oder abwärts zu zählen und beim Erreichen eines bestimmten Grenz- oder Schwellenwertes entsprechende Aktionen, z. B. Entfernen des Paketes aus der Warteschlange, zu triggern. Alternativ könnte auch die Differenz zwischen dem Eintragszeitpunkt (als unveränderlicher Wert) und der aktuellen Zeit (die mitgezählt wird) in einem sich regelmäßig wiederholenden Vergleich als Entscheidungskriterium herangezogen werden.
  • - Die Warteschlangen werden von den Ports nach dem Multi Server Prinzip bedient. Das Multi-Server-Prinzip garantiert eine faire Zuteilung und die optimale Nutzung der verfügbaren Ressourcen auch und gerade bei variablen Paketlängen. Der Scheduling Prozess wird vom ausgehenden Port genau dann angestoßen, wenn dieser frei wird. Dies geschieht aber (unter Kenntnis der (Rest-)Länge des momentan bedienten Pakets) so rechtzeitig, dass die Entscheidung und das nächste auszugebende Paket rechtzeitig mit dem Ende des vorhergehenden zur Ausgabe bereitsteht. Ein weiterer großer Vorteil des Multi-Server-Prinzips besteht auch darin, dass bei Ports unterschiedlicher Bandbreite diese Unterschiede dadurch, dass frei werdende Ports sich quasi selbst neue Daten aus der (den) Queue(s) holen, automatisch bei der Verteilung berücksichtigt werden.
  • - Die Auswahl der als nächstes von einem Port zu bedienenden Warteschlange erfolgt mit Hilfe einer Scheduling-Funktion.
  • - Es werden weitere getrennte Warteschlangen eingesetzt, um zusätzliche, ggf. pro Verteilungsfächer unterschiedliche Kriterien zu berücksichtigen.
  • - Das weitere Kriterium ist ein FIFO (First In First Out).
  • - Die Warteschlangen werden nur bei Bedarf angelegt. Längere Zeit nicht verwendete Queues werden wieder freigegeben (time-out). Hierdurch wird die Zahl der tatsächlich benötigten Warteschlangen reduziert, weil die Zahl der aktuellen Verzweigungsmuster meiste kleiner sein dürfte als die Zahl der aktuellen Kommunikationsbeziehungen.
  • - Die Verkehrsströme werden gezielt ungleichmäßig nach entsprechenden Vorgaben verteilt. Hierzu werden weitere Kriterien in die Verteilungsentscheidung mit einbezogen werden.
  • - Es wird eine adaptive Annäherung an einen erwünschten Zielwert der Verteilung vorgesehen. Bei sich teilweise überlappenden Bündeln ist es möglich, dass die angestrebten gleichmäßigen bzw. vorgegebenen Verteilung nicht optimal bewirkt wird, weil sich Interferenzen zwischen den Bündeln störend auswirken.
    Eine einfache Alternative sieht vor, diese vorausberechnen und bei der Vorgabe der Scheduling-Regeln zu berücksichtigen.
    In komplexen Fällen (z. B. unter sich immer wieder ändernden realen Netzbedingungen), in denen das Netzverhalten und mithin das Verhalten der Komponenten der Erfindung nicht zuverlässig vorausberechnet werden kann, ist es von Vorteil, die Scheduling-Parameter im Betrieb adaptiv anpassen zu können.
    Dies kann entweder im Knoten autonom oder durch einen Anstoß von außen (Vorgabe neuer Regeln) erfolgen.
    Als regulierendes Kriterium kann dazu der Vergleich (resp. die Abweichung) der tatsächlich erzielten Verteilung mit (von) der gewünschten Verteilung dienen. Vorzugsweise sollte dazu im Knoten z. B. die Häufigkeit der Bedienung der verschiedenen Warteschlangen durch die verschiedenen Ports (oder die resultierende Last der Ports pro Verkehrsklasse) gemessen und für die korrigierende Einheit verfügbar gemacht werden.
    Für netzweite Korrekturen können auch Angaben über die mittlere und die maximal aufgetretenen Länge der verschiedenen Warteschlangen von Bedeutung sein.
    Vorzugsweise werden diese Messungen und Verfahren auf Basis der Anzahl der Pakete vorgenommen. Hierbei wird davon auszugehen, dass sich die Auswirkungen von variabel langen Pakete über relativ kurze Zeiträume ausmitteln. Alternativ können diese Betrachtungen und Messungen auch unter Berücksichtigung der individuellen Längen aller einzelnen Pakete angestellt werden.
  • - Zur Realisierung einer bestimmten Lastverteilung innerhalb eines Bündels wird pro Queue für jeden Port eine Zeitinformation gehalten und beim Scheduling mitberücksichtigt. Die Zeitinformation enthält z. B. die letzten Bedienungszeitpunkte der Queue durch die verschiedenen Ports.
    Vorteilhaft kann hiermit z. B. portindividuell ein bestimmter Mindest- und/oder Maximalabstand zwischen zwei Bedienungen eingestellt werden oder die Zeiten könnten relativ zueinander bewertet werden ("Port x kommt erst wieder dran, wenn er doppelt so lange nicht dran war wie Port y"). Solche Zeitinformationen könnten auch zur Bestimmung eines nächsten Ziel-Bedienungszeitpunktes (frühestens, spätestens, optimal) verwendet werden oder gleich als solche abgespeichert werden (Kalender).
    Eine alternative und vergleichsweise einfachere Lösung besteht darin, für jede Queue lediglich mitzuzählen, wie oft sie von welchem Port bedient wurde, und daraus die Entscheidung abzuleiten, ob sie aktuell mit berücksichtigt werden darf. Die Zähler könnten intervallweise oder auch gleitend (Leaky Bucket) zurückgesetzt bzw. dekrementiert werden.
    Auch bei diesen Varianten könnten die Vorgaben und Regeln wie oben beschrieben mit Hilfe von Messergebnissen adaptiv angepasst werden.

The implementation can be done via a simple, regularly clocked round counter (restart in case of overflow). The meter is only relevant locally, synchronization between the nodes is not necessary, the clock accuracy is not subject to any specific requirements, a mixture of nodes with and without this mechanism in the same network is possible at any time. (The mechanism only affects the individual, local way of working within a node.)

• 1. The time criterion is also used to sort out and discard packets that have waited too long for any reason and whose further transmission (within the scope of the associated real-time application) no longer makes sense. This can be done either during scheduling or through a separate, parallel process. Such a process could e.g. B. to set the time information when entering the packet in the queue to a defined initial value, to count it regularly clocked up or down and when a certain limit or threshold value corresponding actions, z. B. Trigger the packet out of the queue. Alternatively, the difference between the entry time (as an unchangeable value) and the current time (which is counted) could be used as a decision criterion in a regularly repeated comparison.
  • - The queues are served by the ports according to the multi server principle. The multi-server principle guarantees fair allocation and optimal use of the available resources, especially with variable packet lengths. The scheduling process is triggered by the outgoing port exactly when it becomes free. However, this happens (knowing the (remaining) length of the package currently being served) in good time so that the decision and the next package to be issued are ready for issue in good time at the end of the previous one. Another great advantage of the multi-server principle is that, for ports with different bandwidths, these differences are automatically taken into account in the distribution by the fact that ports becoming free fetch themselves new data from the queue (s).
  • - The queue to be served next by a port is selected using a scheduling function.
  • - Additional separate queues are used to take additional criteria into account, which may be different for each distribution subject.
  • - The other criterion is a FIFO (First In First Out).
  • - The queues are only created if necessary. Queues that have not been used for a long time are released again (time-out). This reduces the number of queues that are actually required because the number of current branching patterns is likely to be smaller than the number of current communication relationships.
  • - The traffic flows are deliberately distributed unevenly according to the relevant guidelines. For this purpose, further criteria will be included in the distribution decision.
  • - An adaptive approach to a desired target value of the distribution is provided. In the case of bundles which partially overlap, it is possible that the desired uniform or predetermined distribution is not optimally effected because interferences between the bundles have a disruptive effect.
    A simple alternative is to calculate these in advance and take them into account when specifying the scheduling rules.
    In complex cases (e.g. under constantly changing real network conditions) in which the network behavior and therefore the behavior of the components of the invention cannot be reliably predicted, it is advantageous to be able to adaptively adapt the scheduling parameters during operation ,
    This can be done either autonomously in the node or by an external push (new rules).
    The comparison (or the deviation) of the actually achieved distribution with (from) the desired distribution can serve as a regulating criterion. Preferably, this should be done in the node z. B. the frequency of service of the different queues through the different ports (or the resulting load of the ports per traffic class) are measured and made available to the correcting unit.
    For network-wide corrections, information about the average and maximum length of the various queues can also be important.
    These measurements and methods are preferably carried out on the basis of the number of packets. It can be assumed that the effects of variable-length packets are averaged over relatively short periods of time. Alternatively, these considerations and measurements can also be made taking into account the individual lengths of all individual packages.
  • - To implement a specific load distribution within a bundle, time information is kept for each port for each port and is taken into account in the scheduling. The time information contains e.g. B. the last operating times of the queue through the different ports.
    Advantageously, z. For example, a specific minimum and / or maximum distance between two controls can be set for each port, or the times could be evaluated relative to one another ("Port x can only be used again if it has not been used twice as long as Port y"). Such time information could also be used to determine a next target service time (at the earliest, at the latest, optimally) or could be saved as such immediately (calendar).
    An alternative and comparatively simpler solution is to simply count for each queue how often it was served by which port, and to derive from this the decision as to whether it can currently be taken into account. The counters could be reset or decremented at intervals or in a sliding manner (leaky bucket).
    With these variants, too, the specifications and rules could be adapted adaptively as described above with the aid of measurement results.

The invention is illustrated on the basis of the figures Embodiments explained in more detail. It shows:

Fig. 1 shows a distribution fan, which comprises all of a communication link in a network sense usable node and communication paths;

Fig. 2 overlapping distribution fan plurality of communication links in the network of Fig. 1,

Fig. 3 shows the basic form of a solution according to the invention (still without "timestamp") with n = 1 for the most elementary embodiment of the invention, with the direction of the arrow indicating that the port is picking up the next packet to be output, and the dashed triangles for there is an activated scheduling function (from the port),

Fig. 4 shows a solution according to the invention with timestamp information for the packets of the highest priority queue (the timestamp principle can of course also be used for low priority queues up to best effort), which is also evaluated by the scheduler, with the direction of the arrow indicating that the Port picks up the next packet to be output, and the dashed triangles represent the scheduling function activated (from the port),

Fig. 5 shows a solution according to the invention with further auxiliary registers per queue (with auxiliary information for controlling a predetermined traffic distribution for the respective branching group in which they are serving bundle), wherein indicates by the direction of the arrow, that the port to pick up the next to be output packet, and wherein the dashed triangles stand for the activated scheduling function (from the port),

It should be noted here that the branching patterns in FIGS. 1 and 2 only show the structure of the meshing between the nodes. This does not exclude that there may be several parallel physical lines (and even with different bandwidths) on one edge of the graph (ie between two adjacent nodes), which are considered as separate ports in the distribution as in FIGS. 3 to 5 (can).

A particularly nice advantage of the invention is that when using overlapping branching patterns, any combination of mechanisms a) time criterion pro Arbitration delay optimization package, b) Setting a given, "skewed" if necessary Traffic distribution, c) adaptive readjustment to the target Distribution pattern, a very flexible and for almost everyone Adaptations of the network nodes and consequently the network application Network as a whole is made possible.

An example of a specifically uneven distribution of traffic flows according to corresponding specifications by including further criteria in the distribution decision is to occasionally ignore a specific queue for certain ports according to the corresponding rules when scheduling. For example, if the high or highest priority traffic of a branch group is to be distributed in a ratio of 4: 4: 3: 1 to 4 ports in a distribution on 4 ports, the scheduler for the 3rd port will assign the highest priority queue of this branch group for every 4. Scheduling process is not taken into account and for the 4th port it is only included in every 4th scheduling process. The total volume of the highest priority traffic should in any case (clearly) remain below the total capacity provided for this (port 1 + port 2 + x port 3 + R port 4 ) of the bundle.

Of course, asymmetries of the Port bandwidths are compensated. Has z. B. the 4th port compared to the other 3 ports only half the bandwidth, but should still get a twelfth of the total traffic, so the corresponding queue becomes every second time also taken into account.

It should be emphasized that the description of the invention relevant components of the communication network is not to be understood as restrictive. For a relevant One skilled in the art will particularly appreciate that the to understand the terms used functionally and not physically are. Thus, the components can also be partially or entirely in software and / or across multiple physical Facilities can be realized distributed.

Claims (29)

1. network node of a packet-oriented communication network,
whose nodes are connected to one another in accordance with the topology of the network in such a way that a plurality of paths exist between the network nodes,
wherein the network node is contained in at least one distribution subject that is derived from the network topology for a specific communication relationship between a network node designed as a transmitting node and a network node designed as a receiving node and includes all network nodes and paths that are used for a distributed transmission of traffic flows assigned by the communication relationship the communication network can be used sensibly,
from which, at least for the network node, a branching pattern, on which the distribution fan is mapped, is determined, by means of which a bundle of outgoing ports of the network node is defined, over which the traffic flows assigned to the distribution fan are to be distributed, the routes assigned to the ports being at least two run different, neighboring network nodes,
wherein the network node comprises at least one queuing system for the branching pattern, comprising at least one queue, which is provided at least when at least one traffic stream corresponding to the branching pattern is currently being distributed by the network node. 2. Network node according to claim 1, with a separate queuing system for each Branching pattern, the bundle of which is different from the bundles of the rest Branch patterns in at least one port differ. 3. Network node according to one of the preceding claims, with a separate queuing system for at least part of the distribution subjects, provided that they differ are (t, d). 4. Network node according to one of the preceding claims, with a common queue system for at least part of those different distribution subjects, which are mapped to identical branching patterns (i). 5. Network node according to one of the preceding claims, in the at least one queuing system, at least two Queues includes at least one of the two Queues is marked with a priority. 6. Network node according to one of the preceding claims, with time information at least for those in a priority, especially high-priority queue waiting packets which can be read when the packages are in the Queue have been entered or how long they have been are in it. 7. network node according to claim 6, with a value range for the time information, which can be found under Taking into account the transmission time of the shortest possible Packet on the fastest port is chosen so that a Overflow within a given maximum Delay time of a priority packet does not occur. 8. network node according to one of claims 6 or 7, comprising a separate, parallel process for Sorting out and / or discarding packets, their further Transmission not taking into account the time information seems more sensible. 9. Network node according to one of the preceding claims, comprising second time information for at least one of the Queues through which for each port of the queue assigned bundle whose last service time the queue is displayed. 10. arrangement, in particular communication network, comprising at least one network node according to one of the above device claims. 11. Use of a method with the following steps in a network node according to one of the preceding device claims:
an incoming packet is entered in one of the queues, and
a scheduling process for deciding on the output of a next packet from one of the queues is initiated by at least one of the outgoing ports of the network node. 12. Use according to claim 11, where the scheduling process from the port is so timely before being released, the decision is toasted about the next package to be issued and the next package itself is ready in time for the next package of the port immediately following the previous one Package can be delivered. 13. Use according to one of the above Use claims, in the case of multiple filled queues, the are assigned to the port after a first predetermined Criterion is decided which queue is next is operated. 14. Use according to the preceding claim, in the case of at least partial identification of the Queues with priorities in deciding the filled queues with a low priority as long disregarded, we still have at least one filled Queue with one compared to these queues higher priority is to be considered. 15. Use according to one of the two preceding claims, in which the first criterion is designed as one of the following criteria:
Random,
cyclical order,
Size of the bundles,
shortest queue first,
longest queue first,
after weighting the queues,
elapsed time since this queue was last operated, in particular in the form of a FIFO principle in which the least serviced queue is served next, or
Number of ports that are responsible for their operation. 16. Use according to one of the preceding Use claims, where the next package in the case of one with more than one Packet-filled queue according to a second criterion is determined. 17. Use according to the preceding claim, in which the second criterion is designed as a FIFO. 18. Use according to one of the preceding Use claims, where the queues are only created when needed and Released queues that have not been used for a long time become. 19. A method for transmitting traffic flows in a packet-oriented communication network, applied in a network node according to one of the preceding device claims, with the following steps: - incoming packets are entered into a corresponding queue according to their belonging to one of the distribution compartments, and - transmitted via one of the ports of the bundle that is assigned to the distribution subject and whose ports are determined by the branching pattern on which the distribution subject is mapped. 20. Method according to one of the above Method claims, where the traffic flows of a branch group are targeted unevenly across the ports of the assigned bundle be transmitted. 21. Method according to one of the above Method claims, in which one, especially in the case of partially overlapping bundles Approximation to a desired distribution target adaptive adaptation of scheduling rules is effected. 22. The method according to the preceding claim, where the required adjustment is calculated in advance and at the allocation of the scheduling rules is taken into account. 23. The method according to any one of claims 19 to 21, where the scheduling rules while the Communication network can be adapted adaptively. 24. The method according to the preceding claim, in which the adaptation of the scheduling rules in the network node done autonomously and / or by an external push. 25. The method according to any one of claims 23 or 24, where the adjustment taking into account a regulating criterion, in particular a comparison between the distribution actually achieved with the desired Distribution target is effected. 26. The method according to the preceding claim, in which the frequency of the evaluation of the criterion Operation of the different queues and / or the resulting port load per traffic class is measured. 27. The method according to any one of claims 23 to 26, where the mean and / or the for network-wide corrections maximum length of the different queues is measured. 28. The method according to any one of claims 26 or 27, which is measured based on the number of packets. 29. The method according to any one of claims 26 or 27, where taking into account the individual length of the individual packets is measured.