JP3175735B2 - Switch routing method and shooting blanks switch using the same - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a scheduling method for a queuing system having a plurality of queues requesting a plurality of services, and more particularly to a shooting blanks method for achieving a high throughput in an input queue switch. About.

[0002]

BACKGROUND OF THE INVENTION Some applications require the assignment of multiple queues requesting multiple services. The invention relates to an improvement in the scheduling technique in such a queuing system.

[Graph Theory Terms] In order to facilitate understanding of the present invention, some graph theory terms are defined here.

A bipartite graph is, as exemplified in FIG.
A plurality of vertices are 1.1 to 1.4 and 1.5 to 1.8
An undirected graph that can be divided into two sets, such that edge 1.9 is located only between these two sets. A set of queues requesting a set of services can be represented by such a bipartite graph. The sides are
Represents a cue between a first vertex and a second vertex. The presence of an edge between two vertices indicates that the corresponding queue requires the service represented by the second vertex.

[0005] Bipartite matching is a subset of sides such that no vertex is an end point of a plurality of sides. Seeking a match corresponds to selecting one of the permutations (or permutations) between the queue and the requested service.

[0006] Maximum (size) matching is a matching consisting of the maximum possible number of sides. Let n and m be the number of vertices and edges, respectively, of the underlying graph representing the current state. For an N × N system representing N ² queues requesting N services, n = 2N Yes, m
A ≦ N ^2.

[0007] Maximum weight matching refers to matching that also takes into account weights associated with edges in a graph. In the present invention, the weight is equal to the size of the corresponding queue.

[System Model] In order to better understand background knowledge and the present invention, a system model will be described below. As will be apparent to those skilled in the art, the system model is used only to facilitate understanding of the matter described herein and should not be construed as limiting the technical scope of the present invention. .

Although the present invention uses a switching fabric including a crossbar switch having N input ports and N output ports, the present invention employs M input ports and N output ports.
The present invention is also applicable to a switch having three output ports.
This crossbar switch has one time step
One step. ), Any permutation (reordering) of fixed size packets can be routed.
Only one cell can leave the input port or reach the output in one step. Therefore,
Many cells may be queued at the input. Using one queue per input port can cause a head-of-line (HOL) blocking effect and severely limit throughput. For more information about these blocking effects, see Mark
J. Karol, Michael G. Hluchyj, and Samuel P. Morga
n, "Input versus output queuing on a space-divisio
n packet switch ", IEEE Transactions on Communicati
ons, COM-35 (12): 1347-1356, December 1987.

[0010] Various research groups have suggested using N queues per input port (one queue for each of the N output ports).
For example: Thomas E. Anderson, Susan S. Owicki, James B. Sa
xe, and Charles P. Thacker, "High-speed switch sch
eduling for local-area networks ", ACM Transactions
on Computer Systems, 11 (4): 319-352, November 1993 ・ M. Karol, K. Eng, and H. Obara, "Improving the p
erformance of input-queued ATM packet-switching ",
In Proceedings IEEE INFOCOM'92, p.110-115, 1992 ・ Nicholas William McKeown, "Scheduling Algorithms
for Input-Queued Cell Switches ", PhD thesis, Univ
See ersity of California at Berkeley, 1995.

[0011] The techniques described therein (called virtual output queuing (VOQ)) are models used as background knowledge to explain the preferred embodiment of the present invention. Intuitively, according to VOQ, the switch has excellent flexibility in selecting a replacement. FIG. 2 shows V = N = 4.
FIG. 3 is a schematic configuration diagram illustrating an example of an OQ switch.

The arrival rate of packets that need to go from input port i to output port j is represented by α _ij . Rate 50
% Means that, on average, one packet is sent from input port i to output port j every other step. The rate at which packets arrive at input port i is R
_i . This is equal to Σ _j α _ij . Let S _j be the rate at which packets arrive at output port j. This is equal to Σ _i α _ij . In this sum, i represents all input ports, and j represents all output ports. Although not a formal term, a switch is stable when there are no endlessly growing input ports. When R _i , S _j <100% for all i, j, traffic is said to be allowed .

Router Switch and Scheduling An important example of such a queuing system is a router switch used in applications such as the Internet. The present invention uses switches as an important example of a queuing system that can be represented using a bipartite graph. However, it should be noted that the present invention is not limited to a switch, but is applicable to any queuing system that can be expressed using a bipartite graph.

The development of the Internet continues to increase the demand for routers and switches. Switch designs are currently available that can handle tens to hundreds of gigabits per second of broadband and have a large number of input / output ports. In a network system, cells (or packets) arrive at a switch through several input ports. These cells are routed through one or several output ports based on the scheduling method used by the switch. In this specification, the terms packet and cell are interchangeable, and the invention is not limited to any type of cell or packet.

[0015] Broadband switch designers face the problem of providing an efficient technique for scheduling cells through the switching fabric. Currently, from a theoretical point of view, the best approach known is the conventional Maximum Weight Bipartite Matching (MWBM).
Tite Matching). Unless input / output ports need to operate at 100% capacity, MWBM is stable under fairly general probabilistic assumptions, see Tassiulas and Ephremides and McKeown, Ananthara
m, and independently by Walrand. When none of the queues (associated with the input or output ports) grow indefinitely, the scheduling method is said to be stable. For more information about MWBM, see: Leandros Tassiulas and Anthony Ephremides, "Stab
ility properties of constrained queuing systems and
scheduling policies for maximum throughput in mul
tihop radio networks ", IEEE Transactions on Automa
tic Control, 37 (12): 1936-1948, December 1992 ・ Nick McKeown, Venkat Anantharam, and Jean Walran
d, "Achieving 100% throughput in an input-queued s
witch ", In Proceedings IEEE INFOCOM'96, p.296-30
2, San Francisco, CA, March 1996. A traffic pattern when the load on the input / output port is not full is considered to be acceptable .

However, the execution time of the best known method of MWBM is O (N ² log N + Nm) and O (N
^1/2 mlg NC). Here, N is the number of input ports in the switch, m ≦ N ² is the number of non-empty queues, and C is the maximum queue size. See ML Fiedman and RE Tarjan, "Fibonacci heaps for more information on the execution time of the MWBM method.
and their uses in improved network optimization a
lgorithms ", Journal of the ACM, 34: 596-615, 1987 ・ HN Gabow and RE Tarjan," Faster scaling al
gorithms for networkproblems ", SIAM Journal on Com
puting, 18: 1013-1036, 1989. Furthermore, these methods are sequential in nature. Due to the execution time associated with each replacement, a full implementation of the MWBM is impractical. MWBM
Is not likely to be possible in a wideband switch that requires one permutation calculation per cell time. Substitution is equivalent to matching.

On the other hand, maximum (size) bipartite matching (MBM)
There are two main advantages to the scheduling approach based on.

(1) First, the execution time of the best known MBM method is O (N ^1/2 m). For more detailed information on MBM execution times, see: · JE Hopcroft and RM Karp, "An n ^5.5 algorith
m for maximum matching in bipartite graphs ", SIAM
Journal on Computing, 2: 225-231, 1973 ・ EA Dinitz, "Algorithm for solution of a probl
em of maximum flow in networks with power estimati
on ", Soviet Math. Dokl., 11: 1277-1280, 1970 ・ S. Even and RE Tarjan," Network flow and test
ing graph connectivity ", SIAM Journal on Computin
g, 4: 507-518, 1975.

[0019] (2) to a second input of the method, each queue requires only knowledge (N ² bits) as to whether it is empty. In contrast, the MWBM method requires knowledge of the length of each queue (N ² integers, where there is no upper limit on the size of the integers).

However, when compared with MWBM, M
A disadvantage of BM is that MBM is not stable for all allowed traffic patterns. In particular, McKeown et al. Show a simple traffic pattern of N = 3 that is acceptable but not stable under MBM (Nick McKeown, Venkat Anantharam, and Jean
Walrand, "Achieving 100% throughput in an input-qu
eued switch ", In Proceedings IEEE INFOCOM'96, p.2
96-302, San Francisco, CA).

However, it is expected that there will be interesting classes of allowable traffic patterns that will make the MBM stable. This class of traffic, the equilibrium (balanced)
It is a traffic pattern. In the case of a balanced traffic pattern, the load on all I / O ports is even.
The following facts justify the above conjecture.

1. MBM is known to provide 100% utilization for the offline version of the balanced scheduling problem. If each input port needs to send out exactly T cells and each output port needs to receive exactly T cells, route all NT cells with exactly T permutations It is possible. See: G. Birkhoff, "Tres observaciones sobre el algebr
a lineal ", Univ. Nac. Tuceman Rev. Ser. A5, p.147-
150, 1946; ・ Shizhao Li and Nirwan Ansari, "Input-queued swit
ching with QoS guarantees ", In Proceedings IEEE IN
FOCOM'99, volume 3, p.1152-1159, March 1999). Thus, the substitution can be selected using the MBM method T times.

2. Currently, all known simulations of the MBM approach are stable up to 100% load whenever the traffic flow is balanced. Simulations show instability only for unbalanced flows.

The permutation choice problem can be modeled in graph theory terms. A bipartite graph is an undirected graph in which vertices can be divided into two sets as shown in FIG. 1 and each edge is between these two sets. Bipartite graphs can be used to represent switches. In this case, each input port is represented by one set of nodes (nodes, or vertices), and each output port is represented by the other set of nodes. The presence of an edge between two vertices indicates that there is a cell to be routed between the corresponding input and output ports in the switch. The maximal matching is a matching in which it is not possible to add an edge while maintaining the matching.

Most scheduling methods that appear in the literature fall into two categories. Considers queue size (exemplified by MWBM)
And not considered (exemplified by MBM).

In the MWBM approach, a weight value is assigned to an edge of a bipartite graph. In general, the edge from input port i to output port j is assigned a weight equal to the number q _ij of packets queued at input port i and destined for output port j. Then, the maximum weight matching is calculated in the obtained graph. As mentioned above, the MWBM approach provides stability even when traffic patterns are unbalanced. However, the execution time of the best known method is O (N ² log N + Nm) and O
(N ^1/2 log NC). In addition, these methods are sequential in nature (serial, sequential)
Good hardware implementation is difficult. Heuristics that attempt to approximate MWBM have also been proposed, but they also appear to be limited by sequential computation bottlenecks. See: Adisak Mekkittikul and Nick McKeown, "A practica
l scheduling methodto achieve 100% throughput in i
nput-queued switches ", In Proceedings IEEE INFOCO
M'98, p.792-799, San Francisco, CA, March 1998; Anthony C. Kam and Kai-Yeung Siu, "Linear comple
xity methods for bandwidth reservations and delay
guarantees in input-queued switches with no speedu
p ", IEEE International Conference on Network Proto
cols, Austin, Texas, October 1998).

In the MBM approach, the maximum possible number of packets is routed per timeslot. M
Although asymptotically superior to the WBM method (the best known MBM method only has an execution time of O (N ^1/2 m)),
BM is still a difficult method to implement under the speed constraints of broadband switches. Several heuristic approximations have been developed, such as PIM and SLIP.
These are simpler implementations and attempt to approximate maximum size matching. See: Nick McKeown, Martin Izzard, Adisak Mekkittikul,
William Ellersick, and Mark Horowitz, "The Tiny Te
ra: A packet core switch ", In Hot Interconnects I
V, p.161-173, Stanford University, August 1996; ・ Nicholas William McKeown, "Scheduling Algorithms
for Input-Queued Cell Switches ", PhD thesis, Univ
ersity of California at Berkeley, 1995; ・ Thomas E. Anderson, Susan S. Owicki, James B. Sa
xe, and Charles P. Thacker, "High-speed switch sch
eduling for local-area networks ", ACM Transactions
on Computer Systems, 11 (4): 319-352, November 199
3). These heuristic approaches also
It may show instability for unbalanced allowed traffic patterns.

[Scheduling Heuristics for MBM] As will be appreciated by those skilled in the art, MBM
Several heuristics are available for the law. The MBM-related heuristics described here were used in experiments described in Section IVD below.
Used for comprehensive comparative studies of various MBM methods.

PIM : Parallel Iterative Matching (Parallel
Iterative Matching is O (log
n) is a heuristic that converges to a maximal matching (Thomas E. Anderson, Susan S. Owicki, James
B. Saxe, and Charles P. Thacker, "High-speed swit
ch scheduling for local-area networks ", ACM Transa
ctions on Computer Systems, 11 (4): 319-352, Novembe
r 1993). For each iteration, the unmatched input makes a proposal (request) for any output that has cells queued. Unmatched outputs, upon receiving a request, accept one request by making a uniform random selection across all requests. If an input receives multiple permissions, that input randomly selects one of them.

RND : RND heuristics are similar to PIM. The difference is that any unmatched input i makes only one request for a randomly selected output port j such that i has a cell queued.

RR-RND : Round Robin Method RR-
In RND, the scheduling for a given time slot t is determined by a fixed priority schedule. Input i ₁ the highest priority, it makes a randomly selected from among the outputs, such as to have a cell that has been queued. Input i ₂ of the next highest priority, perform this selection from among the outputs that have not been taken, and so on. In the next time slot t + 1, the input port priority is incremented by 1 modulo N. The advantage of this priority scheme is that it can be easily realized with hardware having a pipeline mechanism.

SLIP : SLIP was developed by McKeown and has been extensively analyzed (Nicholas William
McKeown, "Scheduling Algorithms for Input-Queued
CellSwitches ", PhD thesis, University of Californi
a at Berkeley, 1995). Each input and output port has a separate priority wheel (turntable). One input sends requests, one at a time, to every output where it has a queued cell. One output j is based on its own priority wheel,
Choose to accept the request from. If accepted by i, the priority wheel at j is (i
+1) Set to modN. The input port accepts permission based on the priority wheel and turns its own wheel as well.

SHAKEUP : The SHAKEUP method is
It was introduced by the inventor of the present invention in U.S. patent application Ser. SHAKE
UP can be used by itself or in combination with other methods, and has been shown to achieve large performance gains (see Japanese Patent Application No. 11-244648). This method assumes some initial matching on the bipartite graph. The initial matching is generated by other heuristics, for example. The basic idea is that each unmatched vertex in A is forced to match itself, even if it would remove the existing match.
Specifically, the set A is an end point of at least one side.
In each of the unmatched vertices x, one (for example, (x, y)) of the edges serving as the end points is uniformly and randomly selected. If y∈B does not match,
Add (x, y) to the match. y is already an edge (z,
y), z∈A, if (z,
Remove y) from the matching and add (x, y).
Finally, during SHAKEUP, if multiple vertices in A try to match the same vertex in B, one of the competing vertices in A is randomly selected as the winner. Here, it is assumed that SHAKEUP is not implemented in parallel and is executed in a randomly selected order on unmatched vertices in A. SHAKE
The UP returns a match at least as large as the initial match (for detailed experimental and theoretical justifications for SHAKEUP, see the aforementioned US patent application and Japanese Patent Application No. 11-244648).

The heuristics of PIM, RND and SLIP fall into the class of iterative heuristics. Iterative heuristics are not guaranteed to find a maximal match in one run (iteration). Thus, multiple iterations can increase the size of the substitution.

[0035]

As described above, the MBM
And MWBM approaches each have their disadvantages. Therefore, while having the stability of MWBM,
Scheduling methods with low execution time requirements, such as MBM, are desired, especially for implementing broadband switches.

An object of the present invention is to provide a high performance scheduling method for general traffic patterns.

In particular, it is an object of the present invention to achieve excellent stability of the MWBM scheduler and to achieve a simpler MB
An object of the present invention is to provide a switch scheduling method that performs M or MBM-based heuristics.

[0038]

SUMMARY OF THE INVENTION The approach of the present invention is:
Called shooting blanks . The main concept will be described below. The Shooting Blanks method increases traffic through lightly loaded ports by generating dummy packets so that all inputs and outputs have the same arrival rate Λ. After generating the dummy packet, the MBM or MBM related heuristics are applied to calculate the permutation to be routed. Since the lightly loaded ports correspond to the endpoints of the side corresponding to the dummy traffic, the method chooses to include them in the match in a so-called deceptive manner.

The present invention provides a computer-implemented scheduling method, a method used for routing, a shooting blank switch,
It is also realized as a network system using shooting blanks switches.

According to the present invention, there is provided a computer implemented scheduling method for a system having M groups with N input queues as one group. The N input queues are:
Request a different one of the N services. M and N are natural numbers. Thus, the system will have M × N queues. Determining an arrival rate of entries for M groups and N services; identifying one of the M groups and N services having a maximum arrival rate; The M groups and N
Adding a dummy entry to all of the services and groups other than the identified group or services such that the arrival rates of entries for the M groups and the N services are approximately equal; Using a scheduling method that does not consider the length of the input queue.

According to another aspect of the present invention, there is provided a routing method in a switch having a port. The port is one of M input ports and N output ports, and cells are routed between the M input ports and the N output ports. The cell is
Different subsets of ports arrive at different rates. Each of the M input ports has N queues, so that there are a total of M × N queues. The routing method includes identifying one port as having a maximum cell arrival rate and identifying as having a maximum cell arrival rate such that arrival rates for all input and output ports are approximately equal. Adding dummy cells to all input ports and output ports other than the input port or output port, and using a scheduling method that does not consider the queue length, wherein one of the M input ports The arrival rate for a port is the sum of the arrival rates for the N queues of the one input port,
The arrival rate at one output port of the N output ports is the sum of the arrival rates from each of the M input ports to the one output port.

Preferably, the dummy cells determine and fill a dummy flow for the appropriate queue such that arrival rates for all ports are equal;
Generating dummy cells at the determined arrival rate.

Preferably, the filling comprises identifying a port having a maximum arrival rate Λ, and initializing a packet transmission rate γ _ij between an input port i and an output port j to an initial transmission rate α _ij . the method comprising the steps of initializes the input port pointers and output ports pointers, until the arrival rate S _j arrival rate R _i and the output port j of the input port i is smaller than lambda, the input port pointers and the Incrementing the output port pointer and dummy cells for the flow between i and j, if R _i is greater than or equal to S _j , rate β _ij = Λ
−R _i , otherwise β _ij = Λ−S _j , adding, updating the target arrival rate γ _ij to γ _ij + β _ij , and repeating the above steps for all ports. Is performed using a process consisting of

Preferably, the generation includes generating dummy cells at each input port such that the target arrival rate is satisfied, and piggybacking the real cells to the corresponding dummy cells when a real cell is arriving at the input port. And wherein each queue holds a credit value corresponding to the number of dummy cells queued for transmission in the queue, and the scheduler determines when the credit value is greater than or equal to one. And the credit value is decremented by one when the particular queue is selected.

More preferably, the dummy cells are not actually generated, but each input queue holds a number d _ij . This number is equal to the credit value and corresponds to the number of dummy cells queued for transmission from i to j. The number d _ij and the credit value are incremented by a probability γ _ij at each time step.

More preferably, if the real cell is queued for transmission between i and j when the connection between i and j is granted, the real cell is transmitted. .

In another refinement, preferably, no dummy cells are generated and each input queue has a number c
Hold _ij . This number is equal to the credit value.
c _ij and the credit value are floating point numbers corresponding to the number of dummy cells queued for transmission from i to j. The number c _ij and the credit value are incremented at each time step by γ _ij which takes a decimal between 0 and 1.

Preferably, when the connection between i and j is granted, if the real cell is queued for transmission between i and j, the real cell is transmitted.

In another refinement, preferably c _ij ≧ 1
Or if a real cell is in the queue between i and j, the queue informs the scheduler that it can transmit cells between i and j.

A shooting blank switch according to another aspect of the present invention has M + N ports, a rate discriminator, a filling determiner, a dummy cell generator, and a scheduler. The ports comprise M input ports and N output ports, and cells are routed between the M input ports and the N output ports. The cells arrive at different rates for each subset of ports. Each of the M input ports has N queues, so that there are a total of M × N queues. The rate identifier identifies one port as having the highest cell arrival rate. The arrival rate for one of the M input ports is the sum of the arrival rates of the one input port for the N queues, and the arrival rate for one of the N output ports. The arrival rate is the sum of the arrival rates from each of the M input ports to the one output port. The filling determiner determines a dummy flow for the appropriate queue such that the arrival rates for all ports are equal. The dummy cell generator generates a dummy cell at an arrival rate determined by the filling determiner. The scheduler does not consider the queue length.

Preferably, the generator generates a dummy cell at each input port so that the target arrival rate is satisfied, and piggybacks the real cell to the corresponding dummy cell when a real cell arrives at the input port. . Each queue holds a credit value corresponding to the number of dummy cells queued for transmission in the queue, and the scheduler informs about cells in a particular queue when the credit value is greater than one. The credit value is decremented by one when the particular queue is selected.

Preferably, the dummy cells are actually generated.
Instead, each input queue is a few d _ijHold. this
The number is equal to the credit value and the number of transmissions from i to j
To the number of dummy cells that are queued.
The number d_ijIs the target rate γ for each time step_ij
Is incremented with a probability equal to

More preferably, when the connection between i and j is granted, if the real cell is queued for transmission between i and j, the real cell is transmitted. .

In another refinement, preferably, no dummy cells are generated, and each input queue has a number c
Hold _ij . This number is equal to the credit value.
c _ij and the corresponding credit value are floating point numbers corresponding to the number of dummy cells queued for transmission from i to j. The number c _ij is incremented at each time step by γ _ij which takes a decimal number between 0 and 1.

More preferably, when the connection between i and j is granted, if the real cell is queued for transmission between i and j, the real cell is transmitted. .

In another refinement, preferably c _ij ≧ 1
Or if the real cell is in a queue between i and j, the queue informs the scheduler that cells can be transmitted between queues i and j.

According to yet another aspect of the present invention,
A networking system is provided having switches such that a subset of the switches are shooting blanks switches.

[0058]

DETAILED DESCRIPTION OF THE INVENTION [IVA. Shooting Blanks Approach] The preferred embodiment uses a switch to provide a set of flows that makes the arrival rate equal to に お い て at the heaviest loaded port (either input or output). In such a case, the packet is Λ
Arrives at an average value based on

FIG. 3 is a schematic diagram showing an example of an unbalanced flow switch. Here, the port with the heaviest load is the top input port. The Shooting Blanks method increases traffic through lightly loaded ports by generating dummy packets so that all inputs and outputs have the same arrival rate Λ. After generating the dummy packet, the permutation to be routed is calculated by applying MBM or MBM related heuristics. Since the lightly loaded ports often correspond to the endpoints of the side corresponding to the dummy traffic, the method will be tricked into selecting these dummy packets to include in the matching.

The shooting blanks method has three main advantages. (I) It is experimentally shown to be comparable to the superior stability of MWBM while using MBM related methods (see section IVD). (Ii)
The overhead of implementing the Shooting Blanks concept on MBM heuristics is small (see section IVB) and is certainly negligible compared to the complexity of MWBM or MBM. (Ii
i) This approach is modular. That is,
This approach focuses on "pre-processing" the traffic and can be combined with any scheduling method.

[IVB. Implementation of Shooting Blanks] Details of the implementation of the preferred embodiment will be described below. The switch of the preferred embodiment has a quality of service (Qo
S) Operating in the environment, the flow has a contract to determine the allowed bandwidth. This bandwidth is expressed as a rate, and the rate α _ij of the _ij link, ie, the link between input port i and output port j, is the sum of the rates reserved for all customers on that link. . The rate agreed in the contract is the Call Admission Controller (CAC).
ontroller). The network system ensures that no flow exceeds the rate assigned to it. The rate changes only rarely enough to be able to assume that it is constant over a large time window.

The shooting blanks method of the present invention comprises:
It has three components.

(I) In the Fill-Up component, determine a dummy flow for the appropriate input / output pair to ensure that the rates R _i , S _j for all input and output ports are equal. I do.

(Ii) The Generation component is responsible for generating dummy packets at the rate determined by the filling component.

(Iii) In the actual scheduling component, any method selected (preferably M
BM or MBM-related heuristics) can be performed to calculate the permutation.

[IVB. 1 Filling component] Figure 4
FIG. 4 is a schematic diagram showing an example of the effect of filling the traffic in FIG. The filling component is implemented using a greedy procedure. Hereinafter, Λ is the maximum value of the input and output rates. This component attempts to output a new rate γ _ij . γ _ij is initialized to the original rate value α _ij . The greedy procedure proceeds through the input / output ports in order from 1 to N. Two pointers are provided, one for the current input i and the other for the current output j. These pointers are incremented until input and output ports with loads less than $ are found. For both i and j, R _i , S _j <Λ. Even if it is assumed that Λ−R _i <Λ−S _j , generality is not lost.
Add a flow from rate i to j with rate β _ij = Λ-R _i ,
Update the rate γ _ij to γ _ij + β _ij . Move the input pointer to
Move to the next i ′> i such that R _{i ′} <Λ. Repeat this step. This completes the description of the filling component.

It can be seen that the filling component adds at most 2N-1 flows. At each step, it is always possible to add edges to move towards the goal of achieving a uniform load Λ. This is because the sum of the input rates is always equal to the sum of the output rates. Therefore, this method is correct.

[IVB. 2 Generation Component] Realization of the generation component will be described. Let the rate of the port with the heaviest load be Λ. This component is
Dummy flows must be generated at the rate determined by the filling component. To complicate matters further, if the actual flow (eg, from input i) is less than or equal to the rate agreed in the contract, the method of the present invention will satisfy the rate た set by the filling component for i. The rate at which dummy packets are generated must be increased.

To efficiently address the above two requirements for the generating component, a "piggyback" scheme is used. The input i produces a dummy packet destined for the real and dummy flow destinations determined by the filling component, at a full rate of just Λ. All schedulers consider connection ij as 1 or 0 according to the presence or absence of a dummy packet. If a real packet is queued at input i with destination j, it is moved to the beginning of the dummy packet the next time the scheduler grants the ij connection. According to the piggyback method, realization of shooting blanks is expensive policing of real flow
, And using a fixed rate for dummy packet generation on any link (this is an easily implementable method in hardware).

To further improve efficiency, in a more preferred embodiment, dummy packets are not explicitly generated and retained. For every pair of input / output ports (i, j), only the number d _{ij of} dummy packets queued for transmission from i to j is kept. This number is incremented with a probability γ _ij at each time step. If d _ij > 0, the scheduler determines 1 (i.e., the _ij side exists in the bipartite graph); otherwise, it determines 0. The scheduler has an _ij connection (this implies d _ij > 0)
And if q _ij > 0 (meaning that there are real packets queued for transmission), the real packets are routed and d _ij is decremented by one. ij connection is allowed,
If no real packets are queued, simply d
_ij is decremented by one.

FIG. 5 is a further preferred embodiment according to the present invention, and is a schematic diagram showing a counter-based realization of a generation component. In this embodiment, a deterministic approach is used instead of a stochastic approach. This approach avoids the use of very expensive random number generators. In FIG. 5, _let γ _ij be the rate determined by the filling component for the ij connection.
A counter c _ij is provided, and γ is set for each time step.
_ij (which is a decimal number between 0 and 1). The value of the counter is a floating-point number that can take a negative value. As before, the integer q _ij holds the number of real packets. If q _ij > 0 or c _ij ≧ 1, the scheduler determines that there are ij edges in the bipartite graph. If the scheduler allows _ij connection, c _ij is 1
, And if there is a real packet (ie, q _ij > 0), the real packet is routed.

[IVC. 1 Shooting Blanks Switch] FIG. 6 is a schematic configuration diagram showing a preferred embodiment of a shooting blanks switch according to the present invention. The shooting blanks switch shall have M input ports 6.11 to 6.14 and N output ports 6.21 to 6.24. That is, M = 4 and N = 4 in the illustrated switch. After a cell arrives at an input port, it must be routed between the M input ports and the N output ports. These cells arrive at different rates. Each cell is queued in one of several queues (6.711 to 6.714 are some queues). One queue corresponds to a cell being routed from a particular input port to a particular output port. For example, queue 6.
714 is an input port 6.11 to an output port 6.24.
Queue the cells to be routed to. Similarly, queue 6.712 queues cells that are routed from input port 6.11 to output port 6.22. Thus, each input port has N queues corresponding to cells routed from that input port to each of the N output ports. Thus, this switch has M × N queues. In this case, 4
× 4 = 16 queues.

The rate identifier 6.3 identifies the port having the highest arrival rate. The arrival rate for the input port is
The sum of the arrival rates for all N queues at that input port. Similarly, the arrival rate for an output port is the sum of the arrival rates from each of the M input ports to its output port. The fill determiner 6.4 determines a dummy flow for the appropriate queue such that the arrival rates for all ports are equal. The dummy cell generator 6.5 generates a dummy cell at the arrival rate determined by the filling determiner. The scheduler 6.6
Routing is performed based on a heuristic MBM method that does not consider the queue length.

[IVC. 2 Network System Using Shooting Blanks Switch] FIG. 7 shows a preferred example of a network system using shooting blanks switches. The network system 7.1 has several switches. A subset of these switches consists of shooting blanks switches 7.21 to 7.23. The structure of the preferred embodiment of the shooting blank switch is described in section IVC. 1 (FIG. 6).

[IVD. [Experimental results] It should be noted that the experiments described here are merely examples and should not be construed as limiting the technical scope of the present invention. Using these simulation experiments, we show that the Shooting Blanks method gives better results than the conventional scheduling approach without the Shooting Blanks method. It should be emphasized that the Shooting Blanks method can be combined orthogonally with any scheduling heuristics, but will work better with a good method that approximates maximum size matching to balanced inputs. The experiment described here uses an example traffic pattern generated as follows. At any time step, each input port i generates one cell with probability Σ _{j = 1} ^N α _ij . The probability that the destination of the cell generated at input port i is output port j is α _ij / Σ _{j = 1} ^N α _ij .

The order of an input port refers to the number of output ports to which it is matched. Similarly, the order of an output port is the number of input ports to which it is matched. The aforementioned U.S. patent application and Japanese Patent Application No. 11-2446.
Low-order , unbalanced traffic patterns, such as those described in No. 48, are considered exemplary traffic types. This traffic type is a generalization of the traffic pattern described in McKeown et al. McKe
Although the traffic pattern described in own is allowed, the queue grows indefinitely when the scheduling method is MBM (Nick McKeown, Venkat Anantha
ram, and Jean Walrand, "Achieving 100% throughput i
n an input-queued switch ", In Proceedings IEEE INF
OCOM'96, p.296-302, San Francisco, CA, March 199
6, see). For the current simulation, half the input ports and half the output ports support exactly one flow, and the other input and output ports support exactly two flows. Since each flow generates traffic at the same rate, half of the ports are loaded at rate λ and the other half are loaded at 2λ. Using two random permutations to generate a flow at the start of the simulation, N! Each of the individual possibilities is selected with equal probability. For heavily loaded input ports, use both flows defined by permutation. For lightly loaded input ports, only the first permutation is used to define the flow.

Consider a switch of size N = 4 and N = 8. For completeness, we show the actual rate matrix used. The (i, j) component represents the rate from input port i to output port j. _Let α _{N be the} rate matrix for real traffic. The rate matrix for the blank (dummy) is denoted by β _N. The total rate matrix is represented by γ _N = α _N + β _N. Note that α _N is generated as described in the previous paragraph,
β _N is defined in section IVB. Calculated from α _N as described in 1.

When the size N = 4, γ ₄ = α ₄ + β ₄ , that is,

[0079]

(Equation 1) It is.

When the size N = 8, γ ₈ = α ₈ + β ₈ , that is,

[0081]

(Equation 2) It is.

Consider a range of 0 ≦ λ ≦ 1/2. Since the maximum load on any port is 2λ, 2λ is called the simulation rate .

The experiment seeks to empirically evaluate the maximum rate of stability for a particular configuration. However, the configuration includes a switch size, a scheduling method, and a traffic pattern. Although not a formal term, a configuration is said to be stable if the input queue does not grow indefinitely. An empirical definition of stability is as follows: if the sum of the input queue sizes does not grow larger than a certain upper limit Q within a certain number of time steps T, the configuration is declared stable. Is done. Note that increasing Q or decreasing T may have the effect of increasing the experimental stability range of some configurations. Also, increasing either Q or T may increase simulation time. In our experiment, Q = 16N ² cells,
Use T = 256N ² . Stability experiments determine the maximum rate at which a configuration can operate before our stability criteria are violated. Since such stability experiments can be very long, we add extra departure conditions.
Check the total queue size every 1,000 steps. If this sum has not grown since the last check, the system is assumed to be stable.

Table 1 shows the experimental results. Stability rate tested against 4 ports without blank, 4 ports with blank, 8 ports without blank, 8 ports with blank for 9 scheduling methods Have been. The scheduling method is 3
It is grouped into two classes: MBM, high quality MBM heuristics, and low quality MBM heuristics. Heuristic quality refers to how well the MBM behaves with respect to the original traffic before adding the blank. Heuristics with a stability rate lower than Λ are characterized as poor quality for the purposes of this discussion. All iterative methods were performed four times (4 iterations).

The main results are shown in the first row of the table.
For N = 4, MBM is stable up to な し without blank and up to 100% stable with blank. N = 8
In the case of, the use of a blank increases the stability rate from 93% to 1
Increase to 00%.

The next class of scheduling method (this is PIM-SHAKE-4, RND-SHAKE-
4, RRRND-SHAKE-4, and SLIP-SH
AKE-4. ) Are considered high quality heuristics. For this class of heuristics, the use of blanks generally improves. Typical improvements are 93% stability to 9 stability at 4 ports
This is an improvement to 8%, and an improvement from 90% stability to 92% stability at 8 ports. In addition, these methods are N
Performing better than MBM without = 4 blanks can be explained for this example,
Generally not. These heuristics are biased towards nodes with more connected edges in a bipartite graph. This is advantageous for these examples.
This is because a node with two edges needs to support twice the bandwidth.

[0087]

[Table 1]

The last class of scheduling method is:
Relatively low quality MBM heuristics. This class includes PIM-4, RND-4, RR-RND,
And SLIP-4. In general, these methods perform slightly worse with blanks and the stability rate drops from 88% to 86% for all switch sizes. SLIP-4 is an exception. The 100% throughput for SLIP-4 is:
Given our example traffic flow, we can predict. Full synchronization of the priority wheels allows for 100% throughput. But,
SLIP-4 is problematic with balanced traffic in this example. In general, SLIP can show that it is difficult to achieve a high quality solution for many low order traffic patterns (see the aforementioned US patent application and Japanese Patent Application No. 11-244648). The good performance of SLIP without blank in this experiment is an apparent result with simple traffic patterns.

[0089] Other modifications of the present invention will be apparent to those skilled in the art from the foregoing description. Although only some embodiments of the present invention have been specifically described, it is apparent that various modifications can be made without departing from the technical spirit and scope of the present invention.

[0090]

As described in detail above, according to the scheduling method of the present invention, one port is first identified as having the highest cell arrival rate, and then the arrival rates for all ports are substantially equal. As such, add dummy cells to all ports except those identified as having the highest cell arrival rate. In this way, excellent stability of the MWBM (Maximum Weighted Bipartite Matching) scheduler can be achieved and a simpler MBM (Maximum Size Bipartite Matching) can be achieved.
Or MBM-based heuristics can be performed.

[Brief description of the drawings]

FIG. 1 is a schematic diagram for explaining a bipartite graph of a scheduling problem.

FIG. 2 is a schematic configuration diagram of an N = 4 crossbar switch that supports virtual output queuing (VOQ).

FIG. 3 is a schematic diagram illustrating an example of an unbalanced flow switch.

FIG. 4 is a schematic diagram illustrating a new rate of the traffic of FIG. 3 after generation of a dummy flow by a filling component in one embodiment of the present invention.

FIG. 5 is a further preferred embodiment according to the invention,
FIG. 3 is a schematic diagram illustrating a counter-based implementation of a generation component.

FIG. 6 is a schematic block diagram showing a preferred embodiment of a shooting blank switch according to the present invention.

FIG. 7 is a schematic diagram showing a preferred embodiment of a network system using a shooting blank switch according to the present invention.

[Explanation of symbols]

6.11 to 6.14 Input port 6.21 to 6.24 Output port 6.3 Rate discriminator 6.4 Filling determiner 6.5 Dummy cell generator 6.6 Scheduler 6.711 to 6.714 queue 7. 1 Network System 7.21 to 7.23 Shooting Blanks Switch

──────────────────────────────────────────────────の Continuing on the front page (72) Inventor Staflors Corioparos United States, New Jersey 08540 Princeton, 4 Independence Way, NAC Research Institute, Inc. New Jersey 08540 Princeton, 4 Independence Way, NEC Research Institute, Inc. (56) References JP-A-2-164158 (JP, A) JP-A-10-322388 (JP, A) IEEE / ACM Transactions on Networking Vol. 7 No. 2 p188-201 Globecom'99 p1203-1210 (58) Fields investigated (Int. Cl. ⁷ , DB name) H04L 12/56 H04L 12/28

Claims (17)

(57) [Claims] 1. An apparatus comprising: M input ports and N output ports, wherein cells are routed between the M input ports and the N output ports, wherein the cells are different for each subset of ports. The M input ports have N queues each, so that there are a total of M × N queues in the switch routing method: (a) one with the maximum cell arrival rate (B) identifying all input and output ports except the input or output port identified as having the maximum cell arrival rate such that the arrival rates for all input and output ports are approximately equal; Adding a dummy cell to the output port; and (c) using a scheduling method that does not consider the queue length. Has, in the step (a), the arrival rate for one of the input ports of the M input ports, the 1
Is the sum of the arrival rates of the N queues associated with one input port, and the arrival rate for one of the N output ports is one for each of the M input ports. A routing method, which is a sum of arrival rates at output ports. 2. The step (b) comprises: (b.1) determining a dummy flow for an appropriate queue such that arrival rates for all ports are equal, and executing filling; and (b.2). the method according to claim 1, characterized by having the steps of: performing the generation of the dummy cell arrival rate determined the in step (b.1). 3. The filling of the step (b.1) includes: (b.1.1) identifying a port having a maximum arrival rate Λ among the ports; and (b.1.2) input port i. Initializing the packet transmission rate γ _ij between the output ports j to the initial transmission rate α _ij ; (b.1.3) initializing the input port pointer and the output port pointer to 1; .1.4) until the arrival rate S _j arrival rate R _i and the output port j of the input port i is smaller than lambda, incrementing the input ports pointer and said output port pointer, (B.1.5 ) A dummy cell for the flow between i and j, and a rate β _ij = Λ−R when R _i is equal to or greater than S _j.
i , otherwise at rate β _ij = Λ−S _j , adding: (b.1.6) updating target arrival rate γ _ij to γ _ij + β _ij , (b.1) the method according to claim 2, comprising the steps of: .7) for all the ports repeating the step (b.1.4) ~ (b.1.6). 4. The step (b.2) includes generating a dummy cell at each input port so that a target arrival rate is satisfied, and corresponding to the real cell when the real cell arrives at the input port. Implemented by piggybacking on dummy cells, wherein each queue holds a credit value corresponding to the number of dummy cells queued for transmission in the queue, and the scheduler determines when the credit value is greater than one. 3. The method according to claim 2 , wherein the credit value is decremented by 1 when a cell in a particular queue is informed and the particular queue is selected. 5. Instead of actually generating a dummy cell ,
According to the scheduler calculation, it is assumed that a dummy cell exists.
Te, each input queue, said a number equal to the credit value holds the number d _ij corresponding to the number of dummy cells that are queued for transmission from i to j, the number of d _ij and the credit _5. The method of claim 4 , wherein the value is incremented at each time step with a probability [gamma] _ij . 6. When a connection between i and j is granted, if a real cell is queued for transmission between i and j, the real cell is transmitted and d _ij Is 1
6. The method according to claim 5 , wherein the value is decremented only. 7. Instead of actually generating a dummy cell ,
According to the scheduler calculation, it is assumed that a dummy cell exists.
Te, dummy cells each input queue holds the number c _ij, the number c _ij is equal to the credit value, the number c _ij and the corresponding credit values, to be queued for transmission from i to j a floating point number corresponding to the number of the number c _ij and the credit value, claim 4 for each time step, characterized in that it is incremented by gamma _ij taking the fraction between 0 and 1 The described method. 8. When the connection between i and j is granted, if the real cell is queued for transmission between i and j, the real cell is transmitted. The method according to claim 7, characterized in that: 9. If c _ij ≧ 1, or if a real cell exists in a queue between i and j, the queue is
8. The method according to claim 7 , wherein the scheduler is notified that a cell can be transmitted between and j. 10. A shooting blank switch, comprising: M + N ports, a rate discriminator, a filling determiner,
A dummy cell generator and a scheduler, wherein the port comprises M input ports and N output ports, wherein cells are routed between the M input ports and the N output ports; The cells arrive at different rates for each subset of ports, and the M input ports each have N queues, so that there are a total of M × N queues, 1 with the highest cell arrival rate
One of the M input ports.
The arrival rate for one input port is the sum of the arrival rates of the one input port for the N queues, and the arrival rate for one of the N output ports is the M input ports. , The sum of the arrival rates at the one output port from each of the ports, the filling determiner determines a dummy flow for an appropriate queue such that the arrival rates for all ports are equal, and the dummy cell generator includes: A shooting blank switch, wherein dummy cells are generated at an arrival rate determined by the filling determiner, and the scheduler does not consider a queue length. 11. The generator generates a dummy cell at each input port so that a target arrival rate is satisfied, and piggybacks the real cell to a corresponding dummy cell when a real cell is arriving at the input port. Each queue holds a credit value corresponding to the number of dummy cells queued for transmission in the queue, and the scheduler determines, for a cell in a particular queue, when the credit value is greater than one. 11. The shooting blank switch of claim 10 , wherein the credit value is decremented by one when notified and the particular queue is selected. 12. Instead of actually generating a dummy cell ,
In the scheduler calculation, there is a dummy cell
As each input queue holds the number d _ij, the number of d _ij corresponds to the number of dummy cells that are queued for transmission equal to the credit value, from i to j, the number of d _ij is For each time step, the target rate γ _ij
12. The shooting blank switch according to claim 11 , wherein the switch is incremented with a probability equal to: 13. When a connection between i and j is granted, if a real cell is queued for transmission between i and j, the real cell is transmitted. 13. The shooting blank switch according to claim 12, wherein: 14. Instead of actually generating a dummy cell ,
In the scheduler calculation, there is a dummy cell
As a dummy cell each input queue holds the number c _ij, the number c _ij is equal to the credit value, the number c _ij and the corresponding credit values, to be queued for transmission from i to j 12. The shooting method according to claim 11 , wherein the number c _ij is incremented by γ _ij which takes a decimal number between 0 and 1 at each time step. Blanks switch. 15. When a connection between i and j is granted, if a real cell is queued for transmission between i and j, the real cell is transmitted. The shooting blank switch according to claim 14, characterized in that: 16. If c _ij ≧ 1, or if the real cell is i
If there is a queue between j and j, the queue is i and j
The shooting blanks switch according to claim 15 , wherein the scheduler is informed that the cell can be transmitted between and the c _ij is decremented by one. 17. A networking system, wherein a subset of the switches comprises switches such as the shooting blanks switches of claim 10 .