CN1126333C - Smoothly turning and grouping scheduling control method and arrangement - Google Patents

Abstract

The present invention relates to a packet switching method and a device in a computer network, particularly to a smoothly-turning and packet-scheduling control method and a control device. The device is composed of a weight expansion sequence scanner, a weight matrix supervisor and a queue supervisor. The queue supervisor determines a defined column of a weight matrix in the weight matrix supervisor by means of the output value of the weight expansion sequence scanner, and the column is scanned from top to bottom. When a non-zero term of the weight matrix in the weight matrix supervisor is scanned, the queue supervisor provides services for the corresponding flow of the term by means of a packet selector in the queue supervisor; when the column in the weight matrix is scanned, the queue supervisor takes the next value of a weight expansion sequence from the weight expansion sequence scanner, and a corresponding column in the weight matrix supervisor is scanned; the steps are carried out as such. When the weight expansion sequence scanner scans the tail of the sequence, the weight expansion sequence scanner gets back to the head of the sequence to start scanning again.

Description

Horizontal sliding wheel turns packet scheduling control method and control device

One, technical field

The present invention is that a kind of horizontal sliding wheel turns (Smoothed Round Robin is called for short SRR) packet scheduling control method and control device. Belong to the packet-switch technology field in the computer network.

Two, background technology

Along with computer networking technology, the particularly develop rapidly of internet, replaced the networks such as present telephone network, packet dedicated net, telegraph network, cable television network by a kind of single network infrastructure and the needed various business of people are provided is the hot issue that academia and industrial quarters are concerned about always. Because the internet is the network infrastructure that dominates now, so people wish to realize the comprehensive of all-network by the internet. But also can't be competent at this work based on the internet of IP (Internet Protocol) technology at present. This mainly is because present internet can't provide service quality guarantee. Compare with circuit-switched technology, the IP technology is a kind of based on connectionless technology, does not have fixing circuit to connect between end-to-end, and it is active as the elementary cell band of network processes to divide into groups, destination address and relevant information, and the next stop of being transmitted by the grouping of each transit point Independent Decisiveness. Traditional router based on the IP technology only provides to full capacity type service, and therefore order of transmission, bandwidth, time delay, the shake of grouping all can't guarantee. In the initial design of internet, network interconnection has been placed on the first place, and design at that time mainly for be the not strong data service of requirement of real-time.

In order to introduce service quality guarantee in the internet, be necessary demand different business and user are distinguished, and the service of varying level is provided to them. In providing the different switch (or router) of service quality on Packet Based Network, dissimilar business and user realize by the packet scheduling controller. Realization in that multiple not existing packet scheduler is arranged at present has respectively some pluses and minuses. As well distinguishing different business and user based on the packet scheduler of WFQ (Weighted Fair Queueing). Under this scheduler, user's end to end delay character can well be ensured, and this scheduler can treat all users liberally, but the shortcoming of the maximum of this scheduler is its time complexity increase linear along with the increase of user's number. When user's number became very large, it is impracticable that this scheduler just becomes. At least logarithm increase along with the increase of number of users of scheduling time complexity of the scheduler of other that grows up on the basis of WFQ scheduler, this class scheduler all becomes impracticable under high-speed case. The scheduler that another turns based on wheel such as DRR (Deficit Round Robin) although the scheduler time complexity do not increase with the increase of number of users, but they are good not to different users' degree of isolation, can interact between each user, therefore the delay character of this class scheduler is all very poor, can not be used for various types of real time business and interactive service.

Three, summary of the invention

Purpose of the present invention is a kind ofly carried out fine isolation to different users with regard to providing, make all users can obtain corresponding service quality respectively, preferably delay character guarantee can be provided, and the horizontal sliding wheel that makes the end to end propagation delay time of various real-time application be controlled at an acceptable scope turns packet scheduling control method and control device.

Horizontal sliding wheel of the present invention turns the packet scheduling control method and launches the sequence scanning device by weight, the weight matrix manager, and three parts of queue management device form. The output valve that queue management device utilizes weight to launch the sequence scanning device is determined definite row of the weight matrix in the weight matrix manager, and these row are scanned from top to bottom, when non-" 0 " of the weight matrix in scanning the weight matrix manager, queue management device utilizes its minute group selector that comprises that this corresponding stream is served. Behind these row in having scanned weight matrix, queue management device launches the next one value that weighting the sequence scanning device heavily launches sequence from weight again, again corresponding row in the weight matrix manager are scanned, so go on, when weight expansion sequence scanning device scanned the afterbody of sequence, the head that it will come back to sequence restarted scanning. The control method that weight is launched the sequence scanning device is the weight that it stores to be launched sequence carry out scan round. First α that its scanning weight is launched sequence during beginning₁, and then be second α₂, to the last one, and then be α₁, so move in circles. At scanning α_jThe time, weight is launched the sequence scanning device and is sent α to a minute group selector_iValue. The control method of weight manager is that weight matrix is managed, and preserves the binary coding of the weighted value of every flow in weight matrix. New when flowing to when having, the binary coding of the weighted value that weight matrix manager handle is corresponding with this stream joins in the weight matrix, and after stream left, its corresponding binary coding was deleted from weight matrix by the weight matrix manager. Queue management device is comprised of queue management and two parts of minute group selector. Queue management newly to grouping join in the corresponding formation of corresponding stream, and a grouping of leaving away is deleted from corresponding formation. The control method of dividing group selector is when the output link of system is idle, chooses a grouping by level and smooth robin scheduling method and send on the output link from all waiting lists.

Horizontal sliding wheel of the present invention turns the packet scheduling control device by center processor, memory, and NIU, pci bus forms. Master controller adopts Intel X86 cpu controller, and respectively with memory, pci bus links to each other. NIU has several, and links to each other with pci bus respectively.

The present invention has following advantage: SRR packet scheduling control method is unique time complexity with strict O (1) in present all packet scheduling control methods, has again the packet scheduling control method of good short-term fairness and scheduling delay character simultaneously.

The present invention can well isolate different users, makes all users can obtain corresponding service quality respectively; Preferably delay character guarantee can be provided, make the end to end propagation delay time of various real-time application be controlled at an acceptable scope.

Packet scheduler in the past has preferably fairness and scheduling delay character as based on the scheduler of markers scheduling mode (such as WFQ etc.) etc., but the realization of this class scheduler is all very complicated, time complexity is all very high simultaneously, minimum is O (longN), and N is flow number in the system. Under high-speed case, when flow number was very large, in fact their high complexity had limited the practical level of this class algorithm.

The time complexity that all has O (1) the another kind of scheduler that turns based on wheel (such as the scheduler based on DRR), and all fine realizations. But because they can not provide good short-term fairness and delay character, so they can not obtain actual application at the packet network (such as the internet) of supporting in the future multiple business.

Because the present invention can overcome the shortcoming of above-mentioned two class schedulers, and has the advantage of this two classes scheduler, so the present invention should be able to obtain widely application in express network in the future and multiple services network environment.

Four, description of drawings

Fig. 1 is structural representation of the present invention. Its right of possession heavily launches sequence scanning device 1, weight matrix manager 2, queue management device 3. Launch sequence scanning device 1 right of possession in weight and heavily launch sequence 11, weight scanner 12. The corresponding weight matrix 21 of information flow f1 to fn is arranged in weight matrix manager 2. The packet queue qf1 to qfn corresponding with information flow f1 to fn in the weight matrix manager 2 arranged in queue management device 3, and minute group selector 31.

Fig. 2 is control program block diagram of the present invention.

Fig. 3 is the structural representation of control device of the present invention. Controller 5 is wherein arranged, memory 6, NIU 7, pci bus 8.

Fig. 4 is the electrical schematic diagram of the NIU 7 among Fig. 3.

Five, the specific embodiment

Horizontal sliding wheel of the present invention turns the packet scheduling control method and is comprised of following several parts: weight is launched sequence (WSS, Weight Spread Sequence) scanner, weight matrix (WM, Weight Matrix) manager, three parts of queue management device form.

The function that weight is launched sequence (WSS) scanner is under the control of minute group selector the weight in the system to be launched sequence (WSS) to carry out scan round, and to a minute group selector output scanning result. Weight expansion sequence (WSS) is a kind of integer sequence of particular design in the present invention.

The function of WSS sequence scanning device is that the WSS sequence that it stores is scanned. The WSS sequence is the integer sequence that defines as follows:

All WSS sequences form the set of a WSS sequence, SETwss＝{S 1，S 2，...，S n... }, S wherein1Be first WSS sequence, SnBe n WSS sequence, n is SnRank. SnMode with recurrence is defined as follows: S1＝{1} S k＝{α i}＝S k-1，k，S k-1 K＞1,1≤i≤2 whereink-1。

It below is the example of several WSS sequences. S¹：{1} S ²：{1    2    1} S ³：{1    2    1    3    1    2    1} S ⁴：{1    2    1    3    1    2    1    4    1    2    1    3    1    2 1     4} S ⁵：{1    2    1    3    1    2    1    4    1    2    1    3    1    2 1     5    1    2    1    3    1    2    1    4    1    2    1    3    1 2     1}

In the WSS scanner, storing statically rank is K_maxThe WSS sequence. K_maxValue will to distribute the scope of weights by system according to it fixed.

The WSS scanner carries out scan round to the WSS sequence that it stores when work. During beginning, it scans α₁, then be α₂..., untilAnd then be α₁..., so move in circles. At scanning α_iThe time, scanner returns α to a minute group selector_iValue.

The weight matrix manager manages and safeguards the weight matrix in the system. What preserve in the weight matrix is the binary coding of the weighted value (here weight is directly proportional with the resulting bandwidth of stream) of every flow (here stream refers to have identical characteristics, such as common source, destination address, the set of the grouping of port numbers). New when flowing to when having, joined in the weight matrix by the binary coding of the scheduling controller notice weight matrix manager handle weighted value corresponding with this stream; After the open system of wandering about as a refugee, its corresponding binary coding is deleted from weight matrix by the weight matrix manager.

In the present invention, each stream has been assigned with the weight (weight) that a speed of reserving with it (rate) is directly proportional. The set that we establish weight for 1,2,3 ... 2^k-1}. Like this, for different k values, we just can obtain different proportion ranges. As, work as k=16, the partition size of bandwidth is 1bps, the at this moment set of system's bandwidth that can provide be exactly 1bps, 2bps, 3bps ..., 64kbps}; Work as k=32, when the partition size of bandwidth is 1bps, the set of bandwidth be exactly 1bps, 2bps, 3bps ..., 4Gbps}.

For stream flow_f, its weight can represent with following mode: $w_f=\underset{n=0}{\overset{k-1}{\mathrm{\Σ}}}{\mathrm{\α}}_{f,n}.2^n$ α wherein_f，n∈ { 0,1}. We are flow_fWeight w_fBinary coefficient α_f，nFormed flow_fWeight vector (Weight Vector), WV_f＝α _f，(k-1)，α _f，(k-2)，...，α _f，0On the basis of weight vector, the concept that we introduce weight matrix WM is as follows.

For weight matrix WM, its corresponding weights of different row are different. Column₀Corresponding weights are 1, column₁Corresponding weights are 2 ..., column_k-1Corresponding weights are 2^k-1 Therefore, we can stipulate column₀The rank of corresponding weights are 0, column₁The rank of corresponding weights are 1 ..., column_k-1The rank of corresponding weights are k-1. In the WM manager, we couple together non-zero in each row among the WM with a doubly linked list in order to reduce the time complexity of scheduler. When the k row are arranged among the WM, k doubly linked list just arranged in the manager like this.

The present invention by means of WSS sequence and weight matrix WM, designs the SRR dispatching algorithm just. We are introduced the operation principle of SRR packet scheduling controller after queue management device is done concise and to the point description below.

Queue management device comes the grouping in the system is selected by control WSS scanner and weight matrix, with the output order of the grouping of each formation in the decision systems and guarantee that the grouping of each formation obtains the service that they should obtain. Queue management device is finished 2 following functions:

Queue management and minute group selection. Wherein queue management to every flow formed formation safeguard. Namely every flow newly to grouping join in order in the corresponding formation. The effect of minute group selection is to be grouped on the physical link and to be sent when one, and also has in the system in the situation of packet queue, the grouping of selecting the next one to send by it. In the present invention, the selection work of grouping is finished in the assistance that divides group selector to need WSS scanner and weight matrix manager.

Core content of the present invention just is to utilize WSS scanner and weight matrix manager to realize dividing into groups fast selection function.

Queue management device is comprised of queue management and minute these two functions of group selection. Queue management newly to grouping join in the corresponding formation of corresponding stream, and a grouping of leaving away is deleted from corresponding formation.

The function of minute group selection is output link when system during the free time, chooses a grouping and send to output link and get on from all waiting lists. The realization of minute group selection will depend on the WSS scanner in the present invention and the WM matrix is finished choosing next one grouping to be transmitted.

Below the summary horizontal sliding wheel turns packet scheduling controlling party ratio juris.

In the description of the principle of SRR dispatch control method, we adopt following term:

S k	Rank are the WSS sequence of k;
		M	The weight matrix that is formed by the weight of all active flows;
K max	The high-order of used WSS sequence in the SRR algorithm;
		K	The rank value of the currently used WSS sequence that arrives of SRR algorithm;

P c	The index value of the current scanning position of the currently used WSS sequence that arrives, span are [1,2k-1]；
		queue f	From the upper received formed formation of grouping of stream f, it is a fifo queue;
P f	At queuefGrouping on the queue heads;
		L f	P fBlock length (take byte as unit);
w f	The corresponding weight of stream f;
		deficit f	A shaping counter, be used for recorded stream f can epicycle service do not run out of, need bring the credit of next round into;
DL i	In SRR, total KmaxIndividual doubly linked list is used for more effectively representing weight matrix. DLiBe i doubly linked list, 0≤i≤Kmax。
		P dl	Point to the pointer of doubly linked list present node;
L max	The maximum packet that output link allows, the i.e. value of the MTU of output link;
		C	The bandwidth of normalized output link.

We are with three sections following pseudo C codes, Schedule, and Add_flow, Del_flow represent the course of work of SRR scheduler, action of each section coded representation. These three action are asynchronous behavior each other. Each action triggers by some asynchronous events.

Schedule is the major part of SRR scheduler, so long as output link is in busy state, SRR will move Schedule to be come output scheduling is carried out in grouping.

Add_flow is called when new stream joins in the system.

Del_flow calls when stream is deleted from system. Stream can explicitly be deleted by ordering, and also can implicitly be deleted by SRR (after all groupings of stream all have been sent out).

The formalized description of these three action is as described below.

Schedule{ local variable：f，col；/*f，col are the current row，column number of M， respectively*/ P c＝1；P dl＝head k-l->next；/*initialization*/ while(in busy-period){      f＝P dl->fid；      deficit f＝deficit f+L max；      while(deficit f＞0){           if(L f＜＝deficit f){                dequeue(P f)；                send(P f)；                deficit f＝deficit f-L f；                if(queuefis empty){                      Del_flow(f)；                      Break；                }           }else{                 break；           }     }     if(P dl-＞next！＝tail col){           P dl＝P dl-＞next；     }else{ loop：P c＝P c+1；       if(P c＝＝2 k)P c＝l；       col＝k-S k[P c]；/*get the corresponding column number*/       if(DL col is empty)goto loop；       P dl＝head col-＞next；/*points to the first non-zero term of this column*/     }  } }

Add_flow(w){/*w is the weight assigned to this flow*/  local variable：f；  f＝getFlowID()；/*get a flow ID for this flow*/  assign deficit f，queuef；/*deficitf＝0，queuef is empty*/  use w to form a Weight Vector as shown in Definition2  add the vector to the last row of Maxtrix M；  adjust DL 0，DL 1，...，DL Kmax-1；  update k；/*if new columns are added into the M*/ }

Del_flow(f){/*f is the flow ID ofthis flow*/     remove the coressponding row from M     remove deficitf，queuef，     adjust DL 0，DL 1，...，DL Kmax-1；     update k；/*if empty columns are deleted from M*/     P c＝P cmod(2 K)；/*P c should be adjusted if some columns are deleted from M*/ }

Controller is dynamically being adjusted weight matrix M all the time at running.

When new flowing to reaches, SRR will join the weight of this new stream last column of M, if the weight of this new stream is simultaneously $w_f={\mathrm{\Σ}}_{n=0}^{j-1}{\mathrm{\α}}_{f,(j-1)}{2}^n(j>k,{\mathrm{\α}}_{f,(j-1)}=1),$ New row column_j-1，column _j-2，...，column _kTo join among the M, and the weights of WSS sequence will be adjusted into j (being k=i).

When a stream has left SRR, flow therewith one in the corresponding weight matrix be about to deleted, if the row column in the Metzler matrix_k-1，column _k-2，...，column _iBecome sky (value is 0 entirely), these row also will be removed from Metzler matrix, and simultaneously, the rank of employed WSS sequence also will be adjusted into i (being k=i).

S ^kBe that the order of WSS sequence is also dynamically adjusted in SRR by the k rank sequence of the definition defined of WSS sequence, its order is exactly the number of the row of weight matrix. In this algorithm, we have stipulated a maximum, K for the WSS sequence_max According to the character 1 of WSS sequence, K is compared on rank_maxLittle sequence can directly be K from rank all_maxSequence in derive. In the rank of each adjustment WSS sequence, new WSS sequence does not need to be recalculated like this. So that the adjustment on the rank of WSS sequence can not need to pay high time and space cost.

deficit _fBe used for depositing stream f does not have consumption in this service credit, these credits can be used deficit_fBe brought in the next time service. Deficit_fCan be carried out liberally so that dispatch.

In SRR, also has K_maxIndividual doubly linked list DL₀，DL ₁，K，

DL _iBe used for the row column of link weight matrix M_iIn all non-zero. DL_iNode have following data structure:

Struct DL_node{ int flowid; / * is used to refer to the ID * of the stream under this node/int weight; / * is used to refer to the weights * of the row of the weight matrix under this node/struct DL_node*next; / * points to next node */struct DL_node*prev; / * point to previous node */;

Flowid from DL_node can judge which stream this node belongs to, and the weight from DL-node can judge this node belongs to which row in the weight matrix. Next, it is a doubly linked list that the prev pointer is linked into non-zero items all in the row in the weight matrix. Selecting doubly linked list, is the efficient of node being deleted from chained list in order to improve.

When SRR just brought into operation, weight matrix M and all doubly linked lists all were empty; When first flows to when reaching, Add_flow will be activated, and system enters the busy period, and then Schedule will bring into operation; When wandering about as a refugee open system for one, Del_flow will be called; After all stream all left system, system had entered the not busy phase, waited for the arrival in next busy period.

The general Intel X86 chip that is that CPU adopts in the present embodiment, WSS sequence and WM matrix all leave in the memory of system, and the memory of this part can be made of SRAM or DRAM. The queuing of each grouping is also left in the memory of system. Queue management device, WSS scanner and WM manager are realized by the software in the microprocessor on function. In realization, we also can select the CPU of other type to replace X86 CPU. It is pointed out that here hardware design given here is a kind of implementation of SRR. SRR also can at other CPU, realize under bus or the NIU.

Software section of the present invention is divided into following 6 parts: initialization module, NIU driver module, WSS scanning, WM management, queue management module, packet classification module, routing table lookup module.

Wherein initialization module is done some initial setting up after startup, as generating the WSS sequence, initializes related data structures etc.

Other module is by the flow process that is grouped in the scheduler, and their function is described as follows: after NIU (such as Ethernet card) is received a grouping, it will give the NIU driver module this grouping; Then driver module is given the packet classification module grouping and is classified, to distinguish dissimilar groupings; Then the routing module that divides into groups to be given is searched routing table here to determine grouping is from which interface unit output; Behind the interface unit of determining output, the queue management module that corresponding interface unit is just given in grouping. Queue management module is given the NIU driver of lower floor suitable grouping under the acting in conjunction of WSS scan module and WM administration module; The NIU driver is given hardware grouping and is sent to circuit and get on.

Below we illustrate the course of work of the present invention by a concrete example.

If 4 streams are arranged in the packet scheduling controller, from NIU A to NIU E, are designated as A-E respectively; Other flow point is not from B to E, C to E, D to E. Be designated as respectively B-E, C-E, D-E. We are the other called after f of these flow points₁，f ₂，f ₃，f ₄ These flow corresponding reservation speed and are respectively r1=64kbps, r2=256kbps, r3=512kbps, r4=192kbps.

The length of the grouping that all stream produces all is 512 bytes, and all stream all is grouped in queuing (backlogged) when scheduling.

If the bandwidth of the corresponding output link of E is C=2 in the network interface²⁰＝1Mbps。

In this example, the partition size of establishing bandwidth is 64kbps, and therefore, these 4 the corresponding weights of stream are respectively w1=1, w2=4, w3=8, w4=3.

According to the value of these 4 the corresponding weights of stream, f₁，f ₂，f ₃，f ₄Corresponding weight matrix M is: $M=\left[\begin{array}{c}{\mathrm{WV}}_1\\ {\mathrm{WV}}_2\\ {\mathrm{WV}}_3\\ {\mathrm{WV}}_4\end{array}\right]=\left[\begin{array}{cccc}0& 0& 0& 1\\ 0& 1& 0& 0\\ 1& 0& 0& 0\\ 0& 0& 1& 1\end{array}\right]$

Metzler matrix leu from left to right is numbered 3,2,1,0 row. Corresponding WSS sequence is:

                 S ⁴＝{1，2，1，3，1，2，1，4，1，2，1，3，1，2，1}

Turn the packet scheduling control method according to horizontal sliding wheel, scheduler is according to S⁴Sequential value at first scan weight matrix the 3rd row, the 2nd row, and then be 3 row .... 32313230323132 3. therefore that is to say that the order of row of scanning Metzler matrix is, every flow will obtain by following service sequences the service of scheduler:

          f ₃，f ₂，f ₃，f ₄，f ₃，f ₂，f ₃，f ₁，f ₄，f ₃，f ₂，f ₃，f ₄，f ₃，f ₂，f ₃

From top service sequences, we can draw, f₁Serviced 1 time in the wheel scan of WSS, f₂Serviced 4 times, f₃Serviced 8 times, f₄Serviced 3 times. All stream has all obtained corresponding service times by their weights.

Claims (5)

1, a kind of horizontal sliding wheel turns the packet scheduling control method, adopt weight expansion sequence scanning device (1), weight matrix manager (2), (3) three parts of queue management device to control, it is characterized in that at first queue management device (3) utilizes the output valve of weight expansion sequence scanning device (1) to determine definite row of the weight matrix (21) in the weight matrix manager (2), and these row are scanned from top to bottom; When non-" 0 " of the weight matrix in scanning weight matrix manager (2), queue management device (3) utilizes its minute group selector (31) that comprises that this corresponding stream is served; Behind these row in having scanned weight matrix, queue management device (3) launches the next value that weighting the sequence scanning device (1) heavily launches sequence from weight again, scans being listed as accordingly in the weight matrix manager (2) again; So go on, when weight expansion sequence scanning device (1) scanned the afterbody of sequence, the head that it will come back to sequence restarted scanning.

2, horizontal sliding wheel according to claim 1 turns the packet scheduling control method, it is characterized in that it is the weight that it stores to be launched sequence (11) carry out scan round that weight is launched the control method of sequence scanning device (1), at first first α of its scanning weight expansion sequence₁, and then be second α₂, to the last oneAnd then be α₁, so move in circles, at scanning α₁The time, weight is launched sequence scanning device (1) and is sent α to a minute group selector (31)₁Value.

3, horizontal sliding wheel according to claim 1 turns the packet scheduling control method, the control method that it is characterized in that weight manager (2) is that weight matrix (21) is managed, and preserves the binary coding of the weighted value of every flow in weight matrix; New when flowing to when having, weight matrix manager (2) joins the binary coding of the weighted value corresponding with this stream in the weight matrix; After stream left, its corresponding binary coding was deleted from weight matrix by weight matrix manager (2).

4, horizontal sliding wheel according to claim 1 turns the packet scheduling control method, it is characterized in that queue management device (3) is comprised of queue management and (31) two parts of minute group selector, queue management newly to grouping join in the corresponding formation of corresponding stream, and a grouping of leaving away is deleted from corresponding formation; The control method of dividing group selector (31) is when system's output link is idle, chooses a grouping by level and smooth robin scheduling method and send on the output link from all waiting lists.

5, a kind of horizontal sliding wheel turns the packet scheduling control device, it is characterized in that this control device is comprised of center processor (5), memory (6), NIU (7), pci bus (8), master controller (5) adopts Intel X86 cpu controller, and link to each other with memory (6), pci bus (8) respectively, NIU (7) has several, and links to each other with pci bus (8) respectively.

Images