CN1585375A - Dispatching method for virtual destination queuing exchanging system - Google Patents
Dispatching method for virtual destination queuing exchanging system Download PDFInfo
- Publication number
- CN1585375A CN1585375A CNA031536948A CN03153694A CN1585375A CN 1585375 A CN1585375 A CN 1585375A CN A031536948 A CNA031536948 A CN A031536948A CN 03153694 A CN03153694 A CN 03153694A CN 1585375 A CN1585375 A CN 1585375A
- Authority
- CN
- China
- Prior art keywords
- algorithm
- coupling
- input
- rmrr
- pamm
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Abstract
The system comprises consist of N input ends and M destination ends. There is an exchange device from input ends to the destination ends. The exchange device is controlled by using a certain dispatching method. The time of the system is dispersed into discrete time quantum. The package reaching the system queues according to destination in buffer memory area of input end to wait for dispatching and transmission. Its dispatching method includes: the reservation matching algorithm that is found according to threshold matrix of length of line, the method of reservation matching entity dispatching that is generated by algorithm of reservation matching combining with basic poll algorithm the reservation matching algorithm combines with basic poll algorithm, and the process to find parallel and excursion execution of reservation matching and entity dispatching.
Description
Technical field
The present invention is a kind of data packet dispatching to the network communicating system of flowing through, and comprises the algorithm of realization, belongs to computer, network, the communications field.
Background technology
Queuing of most employing input or input queuing and other queuing scheme of lining up and making up in the switching fabric of high speed router now.The switching system that has the input queuing realizes that the key of quick exchange is that a simple and effective dispatching method will be arranged, and realizes the quick transmission of packet from input to the destination end to the input queuing.Below from the model introduction of switching system:
As shown in Figure 1, the switching system of a N * M has N input, and M destination end has a switch from input to the destination end, and this switch is controlled according to certain dispatching method by a scheduler.The time of this system will be dispersed is the discrete time section.The packet that the queuing of virtual destination is meant this system of arrival in the input buffer storage area by the destination transmission of waiting in line to be scheduled.Again be divided into " internal system time period " and " system's external time section " time period.Packet is only at the input that begins to arrive this system of " system's external time section ".The packet method that will be scheduled in the end of each " internal system time period " that arrives input is dispatched their ends with being sent to corresponding target, and each packet will be sent to corresponding target ground end in the end of each " internal system time period ".Each packet is that unit is weighed in the delivery time of internal system with " internal system time period ".As not adding explanation, be assumed to be one " internal system time period " in the delivery time of each packet of internal system." internal system time period " should be less than or equal to " system's external time section ".If one " system's external time section " equals S " internal system time period ", then claim this system to have speed-up ratio (Speed-up) S.
In addition, limited Internet resources only at a plurality of inputs that can connect simultaneously under the situation that following condition is satisfied to the path between the end of a plurality of destinations:
Condition 1: any time, these paths of setting up simultaneously do not have common input and destination end each other.
The packet of the different inputs of this switching system of section arrival is sent to identical destination end possibly at one time, and this has just caused the race problem of input to the destination end.These competitions must be with solving someway.The bag of winning in competition will be sent to their destination end along the path of setting up.Solution is the main task of dispatching method to the race problem of destination end.
Condition 2: in each " system's external time section " lining, each input receives a packet at most.
Condition 3: in each " internal system time period " lining, each input sends a packet at most.
Further, in the beginning of each " internal system time period ", dispatching method is dispatched the packet of each formation.In the end of each " internal system time period ", the packet that the method for being scheduled is chosen will be sent to their destination end then.
Above-described switching system is the abstract model of a series of networks, as has switching system, communication network and other network of input queuing.
The core of switching system is a dispatching method, and the quality of dispatching method directly influences the performance of switching system.The leading indicator of investigating a dispatching method has following several:
A, high percent of pass: a desirable dispatching method is to make percent of pass reach 100%;
B, do not have hungry formation: the formation that dispatching method does not allow an input non-NULL can not get service at endless in the time;
C, fast: in order to adapt to the requirement of high bandwidth, dispatching method can not become the bottleneck of systematic function, and therefore, algorithm must be fast as far as possible makes scheduling decision one time, and promptly the time complexity of algorithm is low as far as possible;
But d parallel computation:, generally require the dispatching method can parallel computation in order to meet the demands;
E, be easy to hardware and realize: but dispatching method must be realized by hardware under existing hardware technology.
Up to now, 99107993.0) and " a kind of controlling schemes of constrained queue system " (Chinese patent application number: 99111196.6) and " in the server scheduling and assign the apparatus and method of queuing client requests " of IBM Corporation (Chinese patent application number: 98125972.3) all belong to this class technology the dispatching method of existing many switching systems to this virtual destination queuing is as by 973 projects " a kind of dispatching method of constrained queue system " of Institute of Software, Chinese Academy of Science's research and development (Chinese patent application number:.In each time period the scheduling of data queued bag is described as usually according to the current load matrix of system and finds out a coupling matrix.The input and the destination end of the corresponding respectively virtual destination queuing switching system that is scheduled of the row and column in load matrix and the coupling matrix.Be made up of for " 0 " or " 1 " element value according to the pairing matrix that a load matrix is found out, according to aforementioned condition 1, the coupling matrix should satisfy:
Condition 4: each row, each row of pairing matrix have one " 1 " at most.
Main representational dispatching method has maximum coupling, basic methods such as poll, PIM and iSLIP.The subject matter of maximum matching process is the complexity height, have hungry formation, and when data flow is inhomogeneous, can not reach very high percent of pass, its main cause is that it does not consider the situation (number of wrapping in the formation and the stand-by period of bag) of wrapping in the formation.Basic polling method is very simple, be easy to hardware realizes, in that evenly to flow down percent of pass very high, but percent of pass is lower under non-uniform flow.The subject matter of PIM method is that " picked at random " is difficult in the express network and realizes.The iSLIP method has solved the problem of PIM method " picked at random ", but they all exist in the lower problem of percent of pass in iteration.Though can reach higher percent of pass through iteration repeatedly, in express network, when input and destination end more (respectively above 16), not have time enough to realize repeatedly iteration.
Summary of the invention
The present invention provides the efficient dispatching method of the reservation coupling RMRR (the reservation coupling that Reserved Match with RoundRobin combines with poll) that combines with poll.This method is used in the scheduling of input according to the packet of the switch of virtual destination queuing.In order to adapt at a high speed and the scheduling of inhomogeneous and equal uniform flow, this method proposition keeps matching algorithm, and proposes to keep " whole dispatching method is mated in the reservation of belt wheel inquiry " that matching algorithm combines with basic polling algorithm.This dispatching method can be widely applied in the switching system of actual high speed, and its application comprises communication network and other network.
The following content in this part is not as adding explanation in addition, and the time period all is meant " internal system time period ".
Dispatching method mainly is made up of main algorithm RMRR and aided algorithm PAMM (the parallel approximate greatly coupling of Parallel ApproximateMaximal Match), and main algorithm and aided algorithm move independently of one another.Aided algorithm PAMM is mainly used to determine to keep coupling, promptly the team that surpasses certain thresholding according to the team leader of virtual destination queuing on each input is determined coupling as much as possible.The RMRR method is set, team in case given over to keep coupling after, in K time period, send K packet (K is a system given in advance greater than 1 positive integer) continuously, just discharge this reservation then and mate.Main algorithm RMRR need call the reservation matching result of PAMM algorithm before each run.Basic poll BRR (the basic poll of Basic Round Robin) is made in the current formation that is not given over to the reservation coupling.As shown in Figure 2, will put x exactly
1, x
2, x
3, x
4And y
1, y
2, y
3, y
4In keep coupling coupling to leaving out, at remaining some intercropping BRR.Specific practice is to establish a pointer on each input.All are not given over to pointer on the input that keeps coupling synchronously at each formation of not mated (comprising sky team) superior displacement (initial position of each pointer should stagger), and each displacement moves on in the next formation of not mated successively.These pointers whenever synchronously are shifted and once just obtain a BRR coupling, and whenever draw the BRR coupling one time, just combine the scheduling decision of an integral body of formation with current reservation coupling, send each input execution to and send.Carry out BRR K time, thereby after obtaining K whole scheduling decision, the reservation coupling that the PAMM algorithm makes new advances again continues operation RMRR and obtains the BRR coupling K time, so repeatedly.RMRR algorithm and PAMM algorithm are offset parallel running by timeslice.The RMRR algorithm calls the last round of matching result of PAMM algorithm at every turn.Whenever obtain a BRR coupling, combine with the last round of reservation coupling that obtains and obtain a whole scheduling decision.Owing to adopt this timeslice skew parallel running mode, the computing time of RMRR algorithm basically with basic polling method equivalence.So this method can provide matching result fast, promptly dispatching method can not become the bottleneck of systematic function.
The RMRR algorithm calls the last round of scheduling result of PAMM algorithm at every turn, whenever obtains a BRR coupling, combines with the last round of reservation coupling that obtains and obtains a whole scheduling decision.This relation is established K 4 the situation of getting and is promptly required PAMM algorithm of operation in 4 time periods as shown in Figure 3, is once kept the result of coupling.Pairing 4 time periods of PAMM2 are T
21, T
22, T
23And T
24, and resulting 4 scheduling result are 4 time period T before it in this 4 times
11, T
12, T
13And T
14In the operation resulting reservation matching result of PAMM algorithm " PAMM1 matching result " combining obtains with current 4 polls " matching result of BRR21, BRR22, BRR23 and BRR24 ".For example, time period T
21The whole decision that constantly obtains " decision-making 21 " is result's " PAMM1 keeps matching result " be combined into that is mated by basic poll of finishing in this time period " matching result of BRR21 " and last reservation.Time period T
22The whole scheduling decision that constantly obtains " decision-making 22 " is by basic poll of finishing in this time period " matching result of BRR22 " and last reservation matching result " PAMM1 keeps matching result " be combined into.Equally, time period T
31The whole scheduling decision that constantly obtains " decision-making 31 " is result's " PAMM2 keeps matching result " be combined into that is mated by basic poll of finishing in this time period " matching result of BRR31 " and last reservation.The time period of the arrow points below the whole scheduling decision is that the result of this scheduling decision carries out the time period that sends the packet in the corresponding coupling formation.For example, the arrow points time period T below the whole scheduling decision " decision-making 21 "
23Be meant that this decision-making " decision-making 21 " will be again every two time periods, promptly at T
23Time period, just be used to send the packet in the corresponding coupling formation.Other situation, and the like.Scheduling decision time migration parallel running mode shown in Figure 3 that Here it is.
The RMRR method:
The purpose of central schedule is that the no conflict on each input of decision sends, because input is lined up by the destination end, so in fact just will determine to send out on each input the packet in which team, so that the packet that sends on each input does not have the destination conflict.This problem can be modeled as the matching problem of bigraph (bipartite graph): regard N input as N summit, be designated as X={x
1, x
2..., x
N, regard M destination as an other M summit, be designated as Y={y
1, y
2..., y
M.If have in i input go to j destination packet promptly on this input j team not empty, then with x
iWith y
jBetween connect the limit, obtain a bigraph (bipartite graph) as shown in Figure 2 like this.
Scheduling problem just is converted into the matching problem of the bigraph (bipartite graph) figure that asks shown in Figure 2.Asking the present best algorithm its time complexity of maximum coupling of bigraph (bipartite graph) is O (N
2.5), asking when coupling, if two parts summit of bigraph (bipartite graph) is unequal, can be on the more virtual summits of the less end of number of vertex, and the two ends number of vertices is equated, but virtual summit does not have the limit to be attached thereto.Therefore, below we might as well suppose that the number of input equates with the number of output, all be N.
The PAMM algorithm is not directly to seek coupling on above-mentioned bigraph (bipartite graph), but the team leader is mated above the team of certain threshold value.The matrix T of a N * N of structure, the i of T is capable, the be 1 as the team leader of j team of i input during with j row assignment above given threshold value, otherwise assignment is 0, title T is team leader's threshold matrix.To being that the bigraph (bipartite graph) of adjacency matrix uses the PAMM algorithm to obtain an approximate greatly coupling with T, to by the formation on the coupling, as " keeping coupling ".The PAMM algorithm is chosen 1 element at first concurrently in every row of matrix T.All be to begin to check one by one from certain element whether each element is 1, stops when being checked through one 1 element in every row.Detected 1 element in each row is kept, and full assignment is 0 on all the other positions, and the gained matrix is designated as T '.In each row of T ', choose 1 element then.All be to begin to check one by one from certain element whether each element is 1 in every row.Stop when being checked through one 1 element.Detected 1 element is kept, and all the other element assignment are 0, just obtain a coupling matrix.
For making the team that surpasses thresholding as much as possible be given over to " keeping coupling ", when using the PAMM subprogram, require PAMM iteration three times to form final reservation coupling at every turn.Promptly after of the T matrix operation once of PAMM algorithm to input, the row, column at the coupling element place found among the T is changed to 0 entirely, then the gained matrix T is continued to carry out each step of PAMM algorithm.Twice of iteration more so continuously.At last three results are merged formation and finally keep coupling.
After the reservation coupling draws, to not being BRR by the formation of the input on the coupling.Obtain BRR coupling at every turn, after current reservation coupling combine scheduling decision of acquisition, send to input.In order to narrate this method, determine following mark:
1) Q
Ij: j destination end formation of i input;
2) | Q
Ij|: formation Q
IjThe number of middle packet, i.e. team leader;
3) t[i] [j]: j the destination end formation Q that reflects i input
IjTeam leader's situation, promptly
Wherein K is team leader's threshold value of presetting;
4) team leader's threshold matrix: T=(t[i] [j]), the t[i of all inputs] [j] value constitutes a 0-1 matrix T=(t[i] [j]), is referred to as team leader's threshold matrix.It is the 0-1 matrix of a N * N, and N is the number of input and destination end;
5) the coupling number of the very big coupling of the last round of PAMM acquisition of n:RMRR algorithm;
6) P[i]: the every scheduling decision of making of taking turns of dispatching algorithm, it is one and sends the instruction array, indicates i input to send P[i] packet in the individual formation, i=0,1,2, Λ, N-1.
The RMRR algorithm steps:
Input: the numbering 0,1,2 of N input, Λ, N-1, the numbering 0,1,2 of virtual destination end formation on each input, Λ, N-1; The matching result of last round of PAMM algorithm.Output: scheduling decision, the instruction array of promptly giving input.
Initialization: pointer i points to i the team (total N pointer, one of each input) of i input.
The 1st step: the matching result that calls last round of PAMM algorithm:
The input numbering of not mated: I={I (0), I (1), Λ, I (N-n-1) }, the numbering of the formation of not mated:
J＝{J(0)，J(1)，Λ，J(N-n-1)}。
The 2nd step: with I[k] the pointer ptr[k of individual input] point to J[(k+p) mod (N-n)] individual team, (k=0,1,2, Λ N-n-1), puts l:=1; (l is used for writing down the number of times of BRR scheduling).0≤p≤N-n-1 is a random number.
The 3rd step: make P[ir]=jr, (r=0,1, Λ, n-1);
P[I[k]]＝J[ptr[k]]，(k＝0，1，2，Λ，N-n-1)；
a
i，P[i]＝1，a
ij＝0，(j＝0，1，2，Λ，N-1，j≠P[i])；
The 4th step: to input output scheduling result:
The 5th step was then called the matching result of new round PAMM algorithm as if l=K, changeed for the 1st step; Otherwise, change the step down.
The 6th step: I[k] pointer of individual input is in formation J (0), J (1), Λ, move after the current location order among the J (N-n-1) one (k=0,1,2, Λ, N-n-1); Put l:=l+1, changeed for the 3rd step.
Called in the RMRR algorithm and asked the PAMM algorithm that keeps coupling, the PAMM algorithm that keeps coupling is asked in narration below.
Ask the PAMM algorithm steps that keeps coupling:
The algorithm input:
1. team leader matrix T=t[i] [j], i=0,1 ..., N-1, j=0,1,2 ..., N-1;
2. the position of current each pointer indication: c[i], i=0,1 ...., N-1, r[j], j=0,1 ..., N-1.
During system initialization, c[i]=i, i=0,1 ..., N-1; R[j]=j, j=0,1 ..., N-1.
Algorithm output:
1. matching result
2. the numbering I={I[0 of input of coupling not], I[1], Λ, I[N-n-1];
3. J={J[0 is not numbered in the formation of coupling], J[1], Λ, J[N-n-1];
4. the initial position of next every column pointer: c[i], i=0,1 ..., N-1; R[j], j=0,1 ..., N-1.
Initialization: m:=0 (m represents iterations)
The 0th step: j=0, i=r[j];
The 1st step: if t[i] [j]=1, changeed for the 4th step;
The 2nd step: if i=r[j]-1, changeed for the 5th step;
The 3rd step: put i=i+1 (mod N), changeed for the 1st step;
The 4th step: put t[i] [j]=0 (i=0,1 ..., j-1, j+1 ..., N-1), r[j]=i+1 (mod N);
The 5th step: if j=N-1 changeed for the 7th step;
The 6th step: put j=j+1 (mod N); I=r[j], changeed for the 1st step;
The 7th step: i=0, j=c[i];
The 8th step: if t[i] [j]=1, changeed for the 11st step;
The 9th step: if j=c[i]+1, changeed for the 12nd step;
The 10th step: put j=j+1 (mod N), changeed for the 8th step;
The 11st step: put M[i]=j, c[i]=j+1 (mod N), t[i] [k]=0, (k=0,1 ..., N-1); T[k] [j]=0, (k=0,1 ..., N-1);
The 12nd step: if i=N-1 changeed for the 14th step;
The 13rd step: put i=i+1 (mod N), j=c[i], changeed for the 8th step;
The 14th step: m++ is if the 0th step was changeed in m＜3;
The 15th step: output matching result
I＝{0，1，2，Λ，N-1}/{i
0，i
1，Λ，i
n-1}，
J＝{0，1，2，Λ，N-1}/{M[i
0]，M[i
1]，Λ，M[i
n-1]}。
Illustrate: if hardware is realized allowing, satisfying under the prerequisite of systematic function, suggestion always is to use new T matrix from the 14th iteration that went on foot for the 0th step at every turn.Promptly, call new T matrix, and the row and column that will wherein obtain the element place of mating is changed to 0 entirely when forwarding for the 0th when step to from the 14th step.Promptly as in the 11st step, be t[i] operation of [j] element, the corresponding operating in the 11st step is left out.
The RMRR method that provides can realize and be applied to the dispatching patcher of any described input by the queuing of virtual destination with any program language.
The beneficial effect of RMRR method is, by using the timeslice offsetting mechanism that keeps matching algorithm and coupling and transmission, do not need each scheduling all to pass through the complicated coupling that calculates, and only do once simple polling operation, greatly reduce the time complexity of scheduling, can be fit to the requirement of express network.The RMRR method has comprised the special treatment method to non-homogeneous data flow, and even data flow has also been provided effective processing, can be fit to all even non-homogeneous data flow simultaneously well.The RMRR dispatching algorithm has 4 advantages of iSLIP equally:
(1) do not have hungry formation: not serviced situation of non-empty queue endless time can not appear in the RMRR algorithm;
(2) quick: the RMRR algorithm is owing to just go out a scheduling result after once basic polling operation, time complexity obviously is better than PIM or iSLIP scheduling algorithm, and the time complexity of RMRR algorithm is almost irrelevant with the number of input and destination end;
(3) being easy to hardware realizes;
(4) high percent of pass: to independent identically distributed uniform data flow, the RMRR algorithm can reach 100% percent of pass, to general non-uniform flow, also can keep very high percent of pass.It can look after the data flow of short formation, can give suitably many send opportunity to long formation again.If singly see matching result once, the matching rate of RMRR algorithm in particular cases may be a little less than the iSLIP algorithm at some, but the time that dispatching office of RMRR algorithm is used be significantly less than the iSLIP algorithm, and irrelevant with the scale of network.Therefore, on the high-speed backbone network, the data volume of passing through in the unit interval is weighed, and the RMRR algorithm is better than other algorithms such as iSLIP greatly.
Description of drawings
Fig. 1 is the virtual destination queuing switching system figure of a NXM;
Fig. 2 is a bigraph (bipartite graph);
Fig. 3 is scheduling decision timeslice skew parallel running mode figure;
Fig. 4 is one 3 * 3 virtual destination queuing switching system figure;
Fig. 5 is input, output queuing routers exchange basic block diagram;
Fig. 6 is the routers exchange structure chart based on the parallel data packet switch.
Embodiment
As shown in Figure 4, one 3 * 3 virtual destination queuing switching system, there are 3 inputs (1), (2), (3) in system, 3 destinations (14), (15), (16), from input to the destination end switch (13) is arranged, this switch is controlled according to a kind of dispatching method by a scheduler.The time of this system will be dispersed is the discrete time section.The packet that arrives this system is waited for the transmission that is scheduled by destination queuing (4), (5), (6), (7), (8), (9), (10), (11), (12) in the input buffer storage.Its dispatching method comprises: ask " belt wheel ask reservation mate whole dispatching method " that keeps matching algorithm and keep that matching algorithm combines with basic polling algorithm and ask the process that keeps coupling and skew execution parallel with whole scheduling according to team leader's threshold matrix.
According to the state that is located at a certain moment shown in Fig. 4 system.Get K=4, then team leader's threshold matrix T of 3 * 3 is:
According to the PAMM algorithm, the reservation coupling of trying to achieve for the first time is:
The not coupling upper port of the 0 expression reservation coupling this time in the wherein reservation coupling in the 2nd row.And the matching result of for the first time basic poll is:
Therefore, complete matching result is for the first time:
And the matching result of for the second time basic poll is:
Therefore, complete matching result is for the second time:
RMRR method available software, firmware, hardware or its make up to be realized.Realize the multiple hardware and software architecture that has of RMRR method, what more than provide is typical a realization.Any to the RMRR method realization and do not require the restriction that realized by this typical case.
Suppose that all length of data package are fixed length.Therefore, it is constant that the team's head bag that each is scheduled away is sent to the needed time of its destination end.If elongated packet then necessarily requires at first to be cut into the fixed-length data bag by system when entering input.
Concrete applicating example
Fig. 5 is input, output queuing routers exchange basic block diagram, and wherein 1,2,3 is the input port of router, and 4 are the input scheduling, and 5 is the exchanging array (crossbar) of router, and 6,7,8 is the destination interface of router.The dispatching algorithm that the present invention provides is used in " the input scheduling " shown in 4 among Fig. 5, and the packet that all input ports of router receive is unified scheduling.The dispatching algorithm that the present invention provides requires the packet that receives to be lined up by destination, as shown in Figure 4 in the embodiment at each input port 1,2,3.According to the situation scheduling of data packet queue in each input port 1,2,3, provided a kind of execution mode of situation in the embodiment.
Fig. 6 is the routers exchange structure chart based on the parallel data packet switch in the high-speed backbone network, it is one 4 * 4 parallel switching fabric of router, the speed of each port is R, wherein 1,2,3,4 is the input port of router, 5,6,7,8 is the packet delivery device (demultiplexor) of corresponding input port, 9,10,11 is the sub-interchanger of the router of N * N, 12,13,14,15 is the packet recombiner (multiplexor) of destination interface, and 16,17,18,19 is the destination interface of router.The dispatching algorithm that the present invention provides can be used in the sub-interchanger of router of the N * N shown in 9,10,11 among Fig. 6, and the packet that the antithetical phrase interchanger receives is dispatched respectively.Concrete application mode is similar to the application of input shown in Figure 5, output queuing routers exchange dispatching algorithm.
Claims (14)
1. a kind of dispatching method of a virtual destination queuing switching system, the switching system that is a N * M has N input, M destination end, the dispatching algorithm that the central scheduler of a switch and this system is arranged from input to the destination end, the time of this system will be dispersed is the discrete time section, the packet that arrives this system by the destination transmission of waiting in line to be scheduled, is characterized in that in the input buffer storage area: ask " belt wheel ask reservation mate whole dispatching method RMRR (the reservation coupling that Reserved Match with RoundRobin combines with poll) " that keeps matching algorithm and keep that matching algorithm combines with basic polling algorithm and ask that to keep coupling parallel and be offset the process of execution with whole scheduling according to team leader's threshold matrix.
2. a kind of dispatching method of a kind of virtual destination queuing switching system according to claim 1, it is characterized in that: the input packet of switching system can be a random length, the packet of random length requires to be cut into by system the packet of fixed length; Also can be fixed length, the packet of fixed length be referred to as packet.
3. a kind of dispatching method of a kind of virtual destination queuing switching system according to claim 1, it is characterized in that: dispatching method RMRR (Reserved Match with Round Robin), comprise three algorithms and a process, ask PAMM (the parallel approximate greatly coupling of the Parallel Approximate Maximal Match) algorithm that keeps coupling, left point between basic polling algorithm BRR (the basic poll of Basic RoundRobin), total algorithm RMRR and the PAMM algorithm of asking the reservation coupling and the process that whole scheduling walks abreast and skew is carried out that the PAMM algorithm combines with the BRR algorithm and forms are to reach the fast dispatch to packet.
4. RMRR algorithm according to claim 3, it is characterized in that: every K time period carried out a PAMM algorithm, it is right to obtain the reservation coupling, K time period calls after being provided with, wherein K is a system given greater than 1 positive integer, promptly the team that team leader on each input is surpassed certain threshold value determines coupling as much as possible, keeps sending the packet in these teams in the certain hour section, is called and keeps coupling.
5. RMRR algorithm according to claim 3, it is characterized in that: per 1 time period, to not having the input and the output of selected reservation coupling, carry out one time the BRR algorithm, per 1 time period, the given reservation matching result of scheduling result that RMRR algorithm invokes PAMM algorithm is last round of whenever obtains a BRR coupling, combines with the last round of reservation coupling that obtains and obtains a whole scheduling decision; Carry out obtaining K whole scheduling decision behind K BRR, the PAMM algorithm has gone out new reservation coupling again, continue operation RMRR and obtain the BRR coupling K time, so repeatedly, RMRR algorithm and PAMM algorithm are offset parallel running by timeslice, the RMRR algorithm calls the last round of scheduling result of PAMM algorithm at every turn, whenever obtains a BRR coupling, combines with the last round of reservation coupling that obtains and obtains a whole scheduling decision.
6. RMRR dispatching method according to claim 3 is characterized in that: reach fast dispatch and transformation task to packet with following seven steps:
The first step: initialization determines the initial poll order of basic polling algorithm;
Second step: call the matching result that keeps matching algorithm PAMM;
The 3rd step: remove the match point that keeps matching method PAMM to after left point between carry out basic polling algorithm, obtain left point between a coupling;
The 4th step: the coupling of the 3rd step with the 4th step combined, obtain a whole matching result;
The 5th step: right according to the coupling of setting up, transmit packet in the respective virtual destination end team of corresponding input to its destination end;
The 6th step: after finishing transmission, adjust the pointer of basic poll;
The 7th step:, then upgrade the matching result that keeps matching algorithm PAMM if second step to the 6th step has carried out K time continuously.
7. PAMM algorithm according to claim 4 is characterized in that: finish to obtain to keep with following 2 steps and mate:
The first step: use t[i] [j] reflection i input j destination end formation Q
IjTeam leader's situation, promptly the definition:
Wherein K is a preset threshold value, | Q
Ij| expression formation Q
IjThe number of middle packet, the t[i of all inputs] 0-1 matrix T of [j] value formation=(t[i] [j]), be referred to as team leader's threshold matrix;
Second step: ask to keep coupling, according to team leader's threshold matrix, finding out an input and destination with any one matching algorithm, to satisfy the coupling of team leader's threshold value requirement right, promptly finds out the input and the destination end of pairing.
8. BRR algorithm according to claim 5, it is characterized in that: on each input, establish a pointer, all are not given over to pointer on the input that keeps coupling synchronously at each formation of not mated (comprising sky team) superior displacement (initial position of each pointer should stagger), each displacement moves on in the next formation of not mated successively, and these pointers whenever synchronously are shifted and once just obtain a BRR coupling.
9. RMRR dispatching method according to claim 6 is characterized in that: each input all carries out independent queuing according to their destination end to its packet that will deliver to each destination end.
10. the first step of RMRR dispatching method according to claim 6 is characterized in that: (1) each input is set up a pointer, total N pointer, and (2) pointer i points to i team of i input.
11. in second step of RMRR dispatching method according to claim 6, it is characterized in that:, call the matching result of last round of reservation matching algorithm PAMM according to the result of PAMM algorithm executed in parallel and the process of skew.
12. in the 6th step of RMRR dispatching method according to claim 6, it is characterized in that: each packet that is mated is sent to its corresponding destination end in the section at one time.
13. the 7th step of RMRR dispatching method according to claim 6, it is characterized in that: the reservation coupling that PAMM algorithm of every operation obtains, to use K time in conjunction with basic poll, keep coupling and just replaced by the new coupling that the PAMM algorithm of parallel running obtains.
14. RMRR dispatching method according to claim 1 is characterized in that: the switching system of this virtual destination queuing can be used to realize a series of network systems, comprises switch, the router of input queuing.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CNB031536948A CN1314245C (en) | 2003-08-22 | 2003-08-22 | Dispatching method for virtual destination queuing exchanging system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CNB031536948A CN1314245C (en) | 2003-08-22 | 2003-08-22 | Dispatching method for virtual destination queuing exchanging system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN1585375A true CN1585375A (en) | 2005-02-23 |
| CN1314245C CN1314245C (en) | 2007-05-02 |
Family
ID=34597810
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CNB031536948A Expired - Fee Related CN1314245C (en) | 2003-08-22 | 2003-08-22 | Dispatching method for virtual destination queuing exchanging system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN1314245C (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN100461759C (en) * | 2005-09-16 | 2009-02-11 | 中国科学院计算技术研究所 | Router exchange structure for supporting uniform exchange and service quality guarantee method |
| CN103188042A (en) * | 2011-12-31 | 2013-07-03 | 重庆重邮信科通信技术有限公司 | Matching method and matching accelerator of Internet protocol (IP) data package |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| GB2334116A (en) * | 1998-02-04 | 1999-08-11 | Ibm | Scheduling and dispatching queued client requests within a server computer |
| US6430181B1 (en) * | 1999-01-15 | 2002-08-06 | Lsi Logic Corporation | Crossbar switch fabric with deterministic maximal scheduling of connection requests with strict scheduling prioritization derived from requested service delays |
| CN1277509A (en) * | 1999-06-09 | 2000-12-20 | 顾钧 | Scheduling method for constrained queue system |
| CN1283048A (en) * | 1999-07-30 | 2001-02-07 | 顾钧 | Control scheme for constrained queue system |
-
2003
- 2003-08-22 CN CNB031536948A patent/CN1314245C/en not_active Expired - Fee Related
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN100461759C (en) * | 2005-09-16 | 2009-02-11 | 中国科学院计算技术研究所 | Router exchange structure for supporting uniform exchange and service quality guarantee method |
| CN103188042A (en) * | 2011-12-31 | 2013-07-03 | 重庆重邮信科通信技术有限公司 | Matching method and matching accelerator of Internet protocol (IP) data package |
| CN103188042B (en) * | 2011-12-31 | 2016-03-30 | 重庆重邮信科通信技术有限公司 | A kind of matching process of IP packet and coupling accelerator |
Also Published As
| Publication number | Publication date |
|---|---|
| CN1314245C (en) | 2007-05-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Ueyoshi et al. | Diana: An end-to-end energy-efficient digital and analog hybrid neural network soc | |
| Page et al. | Framework for task scheduling in heterogeneous distributed computing using genetic algorithms | |
| CN1262944C (en) | Processor structure | |
| US20200192701A1 (en) | Multi-thread systolic array | |
| US9575691B2 (en) | Scheduling user requests in a distributed resource system having a plurality of schedulers and coordinators | |
| CN1518705A (en) | Processor structure | |
| CN1993677A (en) | Task processing scheduling method and device for implementing same | |
| CN109104373B (en) | Method, device and system for processing network congestion | |
| CN1642172A (en) | Method and system for transmit scheduling for multi-layer network interface controller (NIC) operation | |
| CN110516801A (en) | A kind of dynamic reconfigurable convolutional neural networks accelerator architecture of high-throughput | |
| CN111142938B (en) | Task processing method and device for heterogeneous chip and electronic equipment | |
| CN1957329A (en) | Signal processing apparatus | |
| CN111104211A (en) | Task dependency based computation offload method, system, device and medium | |
| KR102653013B1 (en) | Systems and methods for i/o transmissions in queue pair-based nvmeof initiator-target system | |
| CN111143036A (en) | Virtual machine resource scheduling method based on reinforcement learning | |
| CN112799823A (en) | Online dispatching and scheduling method and system for edge computing tasks | |
| US20220101086A1 (en) | Reconfigurable hardware buffer in a neural networks accelerator framework | |
| CN114205353A (en) | Calculation unloading method based on hybrid action space reinforcement learning algorithm | |
| CN1314245C (en) | Dispatching method for virtual destination queuing exchanging system | |
| CN110048966B (en) | Coflow scheduling method for minimizing system overhead based on deadline | |
| Wang et al. | Network-on-interposer design for agile neural-network processor chip customization | |
| CN1471153A (en) | Integrated circuit, ssytem development method, and data processing method | |
| Park et al. | Efficient strong scaling through burst parallel training | |
| CN112685162A (en) | High-efficiency scheduling method, system and medium for heterogeneous computing resources of edge server | |
| CN114429214A (en) | Arithmetic unit, related device and method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| ASS | Succession or assignment of patent right |
Owner name: ZHONGDUN XINAN TECHNOLOGY (JIANGSU) CO., LTD. Free format text: FORMER OWNER: GRADUATE SCHOOL, CHINESE ACADEMY OF SCIENCES Effective date: 20100519 |
|
| C41 | Transfer of patent application or patent right or utility model | ||
| COR | Change of bibliographic data |
Free format text: CORRECT: ADDRESS; FROM: 100039 NO.19(A) YUQUAN ROAD, SHIJINGSHAN DISTRICT, BEIJING TO: 211400 WEST SIDE OF GUOFENG ROAD, YIZHENG ECONOMIC DEVELOPMENT ZONE, JIANGSU PROVINCE |
|
| TR01 | Transfer of patent right |
Effective date of registration: 20100519 Address after: 211400 west side of Guofeng Road, Yizheng Economic Development Zone, Jiangsu Patentee after: Zhongdun Xin'an Technology (Jiangsu) Co., Ltd. Address before: 100039 Shijingshan District, Yuquanlu Road, Beijing No. 19 (a) Patentee before: Postgraduate School, Chinese Academy of Sciences |
|
| C17 | Cessation of patent right | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20070502 Termination date: 20130822 |