Router selects the method for link transmission to mainly contain two kinds in the IP network at present: distance vector (Distance Vector is called for short DV) method and Link State (Link State is called for short LS) method.Computer Networks 3 Tsing-Hua University's publication in 1997
RdEd. with among the OSPF Version 2 of IEEE NetworkWorking Group in 1998 this two kinds of methods have been introduced, its basic ideas are at first by transmitting control message between each routing node, obtain the overall shortest path tree of this node other each nodes in the autonomous system, each node is based on the thought of local optimum then, select optimum route (being shortest path) under certain distance measure as much as possible for own handled message, distance measure generally includes performance index such as link bandwidth, transmission delay, buffer storage length here.The difference of these two kinds of methods is to obtain the method difference of overall shortest path tree: the DV method is directly to utilize the overall shortest path tree of each neighborhood of nodes to obtain the overall shortest path tree of this node, and the overall shortest path tree of LS method to be calculating by this node obtain.
If the network bandwidth is always sufficient, the local optimum can be similar to corresponding to global optimum so, but because network traffic has the sudden of no characteristic dimension, bandwidth resources often may be in rare relatively critical condition, in this case, the router transmission method based on the optimum thought of local can produce a serious problem: route flapping.
According to Fig. 1 the process that adopts the optimum method for routing of certainty local to cause route flapping is described below.Among Fig. 1, left and right two sub-Netcoms cross link a, b and connect, in routing table update cycle t, if link a takes place congestedly owing to paroxysmal Business Stream, the routing node A in the then left subnet will select the c-b link as the route of message in t+1 of following one-period to right subnet according to the optimum method for routing of local.For the same reason, routing node B also can select the b link as the route to the right part subnet in the cycle t+1.Can cause like this in cycle t+1, the communication between the subnet of the left and right sides all concentrates on the link b, and owing to empty the required time of transmission buffer usually much smaller than the routing table update cycle, so link a is in idle condition basically in cycle t+1.And in cycle t+2, counter-rotating has taken place in situation, and the c-a link may as one man be judged to be optimal path by the method for routing of local optimum, thereby cause the link b free time, and link a is congested.So move in circles, do not having under the situation of external disturbance, this route flapping process will stably continue.
Because Deterministic Methods is whole behavioral characteristics of modeling real system accurately generally, in case therefore the state of real system has departed from a priori assumption of this method, this method may cause a unacceptable dominance energy down so, and the system mode of dominance the subject of knowledge and the object of knowledge correspondence may be metastable under some, like this, this Deterministic Methods will be attracted for a long time on following dominance energy, such as producing route flapping.
The object of the present invention is to provide a kind of randomized method for routing,, improve bandwidth utilization to overcome the route flapping problem that the optimum method for routing of existing local is produced.
The present invention realizes randomized method for routing by introduce randomization mechanism in the optimum method for routing of existing local, specifically may further comprise the steps:
One, the routing table of expansion router, and be each node initialization probability P
Opt(S);
Two, new routing table cycle, routing node obtains the overall shortest path tree of this node other all nodes in the system according to basic LS method or DV method;
Three, routing node obtains the overall shortest path tree of each node of being adjacent by direct calculating or network service;
Four, this node of routing node construction is to the set of routes at random of the interior every other node of autonomous system;
Five, judge the current message that needs processing that whether has,, then enter step 6 as if having; If do not have, then wait for;
Six, judge whether the current processing message that needs has priority; If message has priority, then change step 8;
Seven, if message does not have priority, then described routing node is with probability P
Opt(S) the optimum route that obtains by the optimum method for routing of local sends its message of flowing through, with probability 1-P
Opt(S) send its message of flowing through by set of routes at random; Change step 11 then;
Eight, judge the probability P that the high priority message occurs in all messages
hWhether greater than probability P
Opt(S), if probability P
hBe not more than probability P
Opt(S), then change step 10;
Nine, if probability P
hGreater than probability P
Opt(S), then the high priority message with probability
Send by optimum route, with probability
By set of routes transmission at random; The low priority message is all pressed at random with probability 1, and set of routes sends; Change step 11 then;
Ten, the high priority message sends by optimum route with probability 1; The low priority message is with probability
Send by optimum route, with probability
By set of routes transmission at random;
11, judge whether the routing table update cycle finish, if then change step 2; If not, then change step 5.
Randomized method for routing of the present invention can also be adaptive, it is characterized in that: when the routing table update end cycle, dynamically adjust probability P according to formula (4)
Opt(S) value:
In the described step 4, routing node is set up set of routes at random: C (S, D) ≡ { B according to formula (1)
i| L
Opt(S, D)≤l (s, B
i)+L
Opt(B
i, D), L
Opt(S, D)>L
Opt(Bi, D) }-{ B
Opt(S, D) (1) its constraints is (L
Opt(S, D)≤l (s, B
i)+L
Opt(B
i, D)) and Λ (L
Opt(S, D)>L
Opt(B
i, D)) and (2).
Described step 4 also can be to calculate the right set of routes at random in this source-order address a new source-order address that is comprised by heading when occurring first, and storage is applied to the whole route cycle.
Describedly send message by set of routes at random and also further comprise: determine that 1) routing node in the set of routes at random is chosen as the posterior probability of message routing; 2) generate random number; 3) selecting at random, transmit path sends message.
Further describe the present invention below with reference to the accompanying drawings.
Principle about the route flapping generation shown in Figure 1 is described in detail in preamble, repeats no more here.
Fig. 2 has provided the basic procedure of randomizing routing method.Core of the present invention is to introduce randomization thought on the basis of the excellent method in existing local optimization road, both has been applicable to link state routing method, also is applicable to the distance vector method for routing.
At first,, need correspondingly expand routing table, make it comprise optimum route and route two class list items at random for existing router based on basic LS method or DV method.
When each routing node began in each routing table update cycle, utilize basic LS method, in autonomous system, spread (Flood) announcement, comprised the distance of this node and its all neighborhood of nodes in the diffusion announcement that routing node sends, can be used for calculating overall shortest path tree.Routing node utilizes the up-to-date diffusion announcement of other routing nodes, construct the overall topological diagram of autonomous system, and utilize the overall shortest path tree of this node of dijkstra's algorithm construction all nodes in the system, wherein distance measure is generally the weighted average of transmission delay, jumping (Hop) number, queue length or These parameters.Also can adopt basic DV method to obtain the overall shortest path tree of local node.
Then, each routing node obtains the overall shortest path tree of all nodes that are adjacent, and for the node that the CPU bottleneck is arranged, is preferably in and has arranged to obtain by straight-forward network communication after the communication protocol that neighborhood of nodes sends out shortest path tree mutually overall; For the node that bandwidth bottleneck is arranged, the overall shortest path tree of its neighborhood of nodes preferably directly calculates acquisition by local node.
If S is arbitrary routing node in the autonomous system, its adjacent nodal set B (S)={ B
1, B
2..., B
k, adjoint point B
iThe length of pairing adjacent side is l (S, B
i), i=1,2 ..., k.If node S when the destination address of the message of pre-treatment is D, then node S utilize following this message of formula construction the C of set of routes at random (S, D): C (S, D) ≡ { B
i| L
Opt(S, D)≤l (s, B
i)+L
Opt(B
i, D), L
Opt(S, D)>L
Opt(B
i, D) }-{ B
Opt(S, D) (1) wherein, B
Opt(S, D) for not only belonging to shortest path from S to D, but also belonging to the routing node of B (S), if there are a plurality of such nodes, promptly to there being many to estimate identical shortest path, then optional one of them as B
Opt(S, D); L
Opt(X Y) is shortest path length between node X and the node Y.C (S, D) and B
Opt(S is D) once obtaining promptly to be stored and to be applied to the whole routing table update cycle.
In Fig. 1, suppose that node A is a to the best route of right subnet, and node B point is b to the best route of right subnet; Node A is c to the hypo-optimal route of right part subnet, and again through b, and node B also is c to the hypo-optimal route of right part subnet, again through a.Like this, if a destination address sends to node B at the A point via hypo-optimal route c at the message of right part subnet, and node B also may select hypo-optimal route c to send this message, and this process may repeat down arbitrarily, has just constituted so-called route ring.For eliminating the route ring, the method for the invention has further been introduced acyclic constraints:
(L
opt(S,D)≤l(s,B
i)+L
opt(B
i,D))Λ(L
opt(S,D)>L
opt(B
i,D))????(2)。
After the set of routes construction was finished at random, routing node need be judged the current message that needs processing that whether has, if do not have, then waits for.If the message that needs to handle does not have priority, routing node S is with probability P so
Opt(S) by optimum route B
Opt(S D) sends message; With probability P
Rnd(S) ≡ 1-P
Opt(S) send message according to set of routes at random.If message has priority, and the probability that message is endowed high priority is P
h, then the routing rule of routing node S transmission message need be made following modification:
1) if P
h≤ P
Opt, then the high priority message sends by optimum route with probability 1; The low priority message is with probability
Send by optimum route, with probability
By set of routes C (S, D) transmission at random.
2) if P
h>P
Opt, then the high priority message will be with probability
Press optimal path and send, with probability
By set of routes C (S, D) transmission at random; The low priority message is pressed set of routes C (S, D) transmission at random with probability 1.
For example, if P
Opt=0.4, P
h=0.25, then all high priority messages will send by optimum route with probability 1, and the low priority message will be with probability
Send by optimum route, send at random with probability 0.8.
When sending at random, determine C (S, D) the routing node C in according to formula (3)
jBe chosen as the posterior probability of message routing:
Wherein, r be used for adjusting path distance is estimated and the selecteed probability in path between relation, this posteriority probability is used to select random walk.
For the simulation chance mechanism, need random number generation source to generate random number.Random number can be generated by the pseudorandom number generator of routing node, also can directly utilize some field in the message as random information source.If a routing node does not satisfy the random walk of acyclic routing rule, its all messages of then flowing through are forced by optimizing route B
Opt(S D) sends.
Fig. 3 is the flow chart about adaptive randomizing routing method.The difference of non-self-adapting method and adaptive approach is key parameter P
Opt(S) setting.
The thinking of self-adapting random method for routing is by dynamic adjustment probability P
Opt(S) numerical value is realized the load balance of optimizing.It is to adjust tactful effect decision by previous step that adaptivity is mainly reflected in current adjustment strategy.Specifically, when t routing table update end cycle, the self-adapting random method for routing according to each output link length of routing node and variation, calculate probability of following one-period by formula (4)
Value.
Wherein
Threshold function table T guarantees that the absolute value of each adjustment amount is not less than threshold value h, and the null value Δ D (S) that can avoid chance to occur makes P thereafter
Opt(S) all make zero; Normalized function N is used for standard
Span, make its not can owing to
The randomness great fluctuation process and acute variation.
Probability P
Opt(S) after the adjustment, remaining flow process is identical with the randomizing routing method of non-self-adapting.
Obviously, work as P
Opt(S)=1 o'clock, the present invention and basic Link State route or distance vector route local best practice are equivalent.In the self-adapting random method, might as well be with P
Opt(S) initial value is taken as 1, thereafter P
Opt(S) can regulate adaptively by control strategy.
In order to verify effect of the present invention, the present invention has been carried out emulation.Network topology structure shown in Figure 1 has been adopted in emulation, and relates to 3 class objects: link, node and message.Link is responsible for the transmission of message, and node also is the generation source of new message except being responsible for routing forwarding, and business model has been used the Poisson process and the paroxysmal mark differential noise process of non sudden respectively.
The parameter of each message object mainly comprises generation time, death time, abandons sign, source address, order address and priority.All message parameters all are initialized as-1, dynamically are provided with and revise in simulation process.Priority list is understood the service type that message requires, and priority of the present invention is only set two kinds of " height ", " low ", has represented the high quality-of-service requirement respectively or has hanged down quality of service requirement.In all simulation processes, each newly-generated message is with probability P
b=0.25 is endowed high priority, with probability P
l=0.75 is endowed low priority.
Fig. 4 shows under the situation that service source is a Poisson process, the probability P of artificial network key event A and B when utilizing the method for the invention to carry out route
OptAnd source, destination address be the relation between the average transfer delay of striding the subnet message (Average Delay) of same subnet not, and wherein transverse axis is represented probability P
Opt, the longitudinal axis is represented average transfer delay.Three subgraphs in upper, middle and lower show the average transfer delay (not considering priority) of all messages, the average transfer delay of high priority message and the average transfer delay of low priority message respectively.As shown in Figure 4, be under the situation of non sudden Poisson model at service source, introduce the average transfer delay that the randomization route has reduced message significantly.
Below table 1 provided all kinds of messages entirely at random route (corresponding to P
Opt=0), basic LS route is (corresponding to P
Opt=1) and optimize at random route (corresponding to optimizing P
OptValue) average transfer delay under.
All kinds of messages of table 1 are at P
OptAverage transfer delay under the representative value of parameter relatively
The message classification | ????P
opt=0
| ???P
opt=1
| Optimize P
opt |
Whole messages | ????70.9 | ????48.0 | ????19.4 |
The high priority message | ????72.3 | ????46.8 | ????3.48 |
The low priority message | ????70.5 | ????48.3 | ????21.7 |
Figure 5 shows that the probability P when service source changes paroxysmal mark differential noise model into
OptAnd the relation between the average transfer delay (Average Delay).On qualitative meaning, its result is consistent with Fig. 4.But there are two important differences in mark differentiated services source and Poisson traffic source: at first, the bursts of traffic source can reduce the absolute performance of system to a certain extent, and this is because sudden utilization more unbalanced the causing on time domain that makes link bandwidth of Business Stream; Secondly, with respect to the Poisson traffic source, the bursts of traffic source makes the average transfer delay of (owning) message and the average transfer delay and the P of low priority message
OptBetween non-linear relation very not remarkable, but this is the cost that increases to the message dropping rate: preceding 10 minimum values of the average Loss Rate of above-mentioned two class messages average is respectively 2.8 and 2.9 times of corresponding index under the Poisson traffic source.
Method for routing of the present invention adopts randomized load-balancing mechanism, has avoided route flapping effectively, realizes the utilization of right idle suboptimum link, has improved network overall transfer performance.Has significant non-linear dependence between the overall transfer performance of IP network and the degree of randomization of method for routing.This non-linear relation for the non-bursty traffic source still be the bursts of traffic source all be significant.The self-adapting random method for routing is to optimum route probability P
OptSelf adaptation adjustment strategy realized acyclic self-adapting random route effectively, improved the overall transfer performance of network significantly.In addition, with optimum route and at random the message that requires of route and different service quality be mapped naturally, can improve the transmission performance of the message of high quality-of-service requirement more significantly, for the solution of service quality provides a kind of new approaches.