JP5888684B2 - Flow aggregation apparatus, method, and program - Google Patents

Description

  The present invention relates to a flow aggregation apparatus, method, and program, and in particular, it is possible to set a route by traffic engineering (TE) in units of a macro flow (MF) that is a set of arbitrary traffic flows between the same departure and arrival (OD) node pairs. The present invention relates to a flow aggregating apparatus, method and program for generating an MF with a small amount of traffic fluctuation in order to suppress the number of MFs in order to suppress TE control cost and to improve the stability of a route.

  The Internet does not manage the state of individual flows in the network and performs routing processing based on the received address. Within a single autonomous system (AS) managed by each ISP (Internet Services Provider), packets are forwarded on a route that has a minimum sum of link weights fixedly set by OSPF (Open Shortest Path First). . In addition, since the transmission host autonomously determines the transfer rate without performing flow admission control, link congestion may occur in principle. Therefore, in order to maintain appropriate communication quality, designing and managing the NW so that the load on all links constituting the network (NW) is below an appropriate level is one of the major issues for ISPs. The link load always changes due to changes in the route caused by fluctuations in AC traffic demand and various faults. For this reason, many traffic engineering (TE) studies have been made to dynamically set a path through which a packet flows so as to actively use a link with a low usage rate (see Non-Patent Document 1, for example). . However, many of these methods are based on the assumption that OSPF is used, and route control is performed by setting link weights appropriately. That is, since the route is set in units of the target NW address, the granularity of the route control is rough, and it is expected that it is difficult to realize a high link load leveling effect.

  On the other hand, in recent years, attention has been paid to SDN (Software Defined Network) technology represented by OpenFlow. In OpenFlow, it is possible to define a flow by arbitrarily combining header information such as an incoming / outgoing IP address, an incoming / outgoing port number, and a protocol number, and to control a route in units of flows. By using SDN, it is possible to easily control the path with fine granularity in units of flow, which is difficult with the conventional technology. In OpenFlow, the controller centrally controls each node in the NW, so that the route setting policy can be changed flexibly, and dynamic route control in a short time period can be realized. For example, Hong et al. And Jain et al. Have proposed a TE method that switches a route in a short time with a fine granularity by using SDN for inter-data center NW (see, for example, Non-Patent Document 2).

  By using SDN in this way, path control in finer granularity is possible, and a high link load leveling effect is expected. However, on the other hand, each node or controller needs to maintain the path status of each path control unit. For example, the flow defined by the 5-tuple of calling IP address, calling port number, and protocol type is the granularity of routing control. In this case, it is expected that the number of states to be managed becomes enormous and it is difficult to apply to a large-scale network. In addition, TE is performed based on future traffic prediction values, but the time fluctuation of the traffic of individual flows is larger than when a route is set in units of the target NW address, so the performance of routing control is degraded due to mispredictions. Is concerned. However, on the other hand, when a route is set with coarse granularity in units of the target NW address as in the conventional route control method using the link weight of OSPF, a large amount of traffic is subject to one route control. There is a concern that the load leveling effect of control will be reduced. In addition, the detour destination link of the route becomes a new high load, and the route is changed again. As a result, the route may become unstable.

  Therefore, it is desirable to implement TE with a granularity that is larger than the individual flow and smaller than the same target NW address. That is, it is effective to define a macro flow (MF) in which some flows are aggregated from a flow set having the same arrival / departure NW address, and to perform path control in units of MF.

A. Elwalid, C. Jin, S. Low, and I. Widjaja, "MATE: MPLS Adaptive Traffic Engineering," IEEE INFOCOM 2001. C. Y. Hong, et al., "Achieving high utilization with softwaredriven WAN," ACM SIGCOMM 2013. S. Jain, et al., "B4: Experience with a globally deployed software defined WAN," ACM SIGCOMM 2013. Y. Hu, D. M. Chiu, and J. C. S. Lui, "Entropy based adaptive flow aggregation," IEEE / ACM Transactions on Networking, 17 (3), pp. 698-711, 2009. Y. Zhang, et al., "Online identification of hierarchical heavy hitters: algorithms, evaluation, and applications," ACM IMC 2004.

  The following are the requirements that should be taken into account for the differences that generate the MF.

  (1) Number of generated MFs M: Routes are set for each TE control target based on the traffic volume. Therefore, it is necessary to measure the traffic at each time point for each TE control target and to predict the future traffic volume if necessary. In addition, since the amount of calculation required when optimally designing a route increases rapidly as the number of TE control targets increases, it is desirable that the number M of generated MFs is small.

(2) MF traffic fluctuation amount V _x : Since the traffic amount of each flow fluctuates, the traffic amount of MF that aggregates a plurality of flows also fluctuates. The TE control target is determined based on the traffic volume rounded at a fixed value, such as the traffic volume at a certain point in time for each path change target and the predicted value of the average traffic volume for a certain period, thus increasing the TE effect. the traffic variation amount V _x of the MFx is smaller is desirable.

  (3) MF maximum traffic volume Q: When the route of a certain MF is changed by TE, the load on the link through which this MF newly passes increases. However, if the maximum traffic volume of the MF is large, The load on the link increases greatly. As a result, there is a possibility that congestion will newly occur in the route change destination link, this MF will be changed to another route again, and the route will become unstable. In order to stably change the route by TE, it is desirable that the maximum traffic volume Q of MF is small.

  There have been some studies on the construction method of MF that aggregates multiple flows. For example, for the purpose of suppressing memory consumption in flow measurement, a method has been proposed for efficiently finding a flow aggregation pattern with a large traffic volume by hierarchically searching IP header information (for example, non-patent literature) 3). A similar flow aggregation method has been proposed for the purpose of finding a flow with a large traffic volume (see, for example, Non-Patent Document 4). However, these methods do not assume that the MF is configured as a TE control target, and cannot satisfy the above three requirements.

As the number of generated MFs is decreased and a large number of flows are aggregated into MFs, the traffic fluctuation amount V _x of the generated MFs decreases due to the statistical multiplexing effect. However, on the other hand, an increase in the maximum traffic volume Q of the MF is expected. Therefore, there is a negative correlation between Q and M and V _x, and it is difficult to optimize all three design goals.

  The present invention has been made in view of the above points, and in order to perform traffic engineering for operating all links of a network with an efficiency of an appropriate level or higher, a macro flow (MF) It is an object of the present invention to provide a flow aggregation device, method and program that can be aggregated into units and that can calculate an optimum route.

According to one aspect, in a network in which a route can be set by traffic engineering (TE) in units of a macro flow (MF) that is a set of arbitrary traffic flows between a pair of identical departure and arrival (OD) nodes, the flow of OD network addresses A flow aggregation device that aggregates in units of MF,
Parameter input means for dividing the number of generated MFs and time into time slots (TS) with a fixed period T and inputting parameters including the generated traffic amount R _{f, t} of each flow f in each TSt in the evaluation period 1 ≦ t ≦ L When,
MF generating means for suppressing the number of MFs in order to suppress the TE control cost, and generating MFs with a small traffic fluctuation amount in order to increase the route stability, using the parameters,
Have
The MF generating means is
A flow aggregating apparatus including means for classifying F flows existing between each OD node pair into M MFs using the generated traffic amount R _{f, t} of each flow f is provided.

  According to one aspect, in a network in which a route can be set by traffic engineering (TE) in units of a macro flow (MF) that is a set of arbitrary traffic flows between the same departure and arrival (OD) node pairs, the TE control cost is suppressed. Therefore, in order to suppress the number of MFs and increase the route stability, there is an effect that an MF with a small traffic fluctuation amount can be generated.

The structural example of the flow aggregation apparatus in one embodiment of this invention. Flow to minimize M and V _x based on greedy algorithm in the first embodiment of the present invention. The flowchart of the macroflow production | generation using the clustering in the 2nd Embodiment of this invention. The flowchart of the spherical k-means method in the 2nd Embodiment of this invention. The example which extracts a part of traffic from each cluster in the 2nd Embodiment of this invention, and aggregates it in each MF. The figure which shows the average value of the traffic variation coefficient of MF produced | generated between Kansas City and Los Angeles. The figure which shows the average value of the MF traffic variation coefficient produced | generated between Chicago and Washington. The figure which shows the cumulative distribution of the traffic variation coefficient of each MF. The figure which shows the maximum peak traffic amount of MF produced | generated between Kansas City and Los Angeles. The figure which shows the maximum peak traffic amount of MF produced | generated between Chicago and Washington. The figure which shows the time series of the maximum link load. The figure which shows the time series of the standard deviation of a link load.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 shows a configuration example of a flow aggregation device according to an embodiment of the present invention.

  A flow aggregation device 100 shown in the figure includes a parameter input unit 101, a microflow (MF) generation unit 102, and an MF output unit 103.

The parameter input unit 101 generates the number of generated MFs M required for the MF design, the generated traffic amount R _{f, t} of each flow f in each time slot (TS) t within the evaluation period, and MF generation using clustering Enter the number of clusters K and weight parameter β required for the modulo.

  The MF generation unit 102 generates M MFs based on the parameter value input from the parameter input unit 101, and the generated MF is output by the MF output unit 103.

  Next, a flow aggregation method for route control according to the embodiment of the present invention will be described.

Consider that the F flows existing between each pair of departure and arrival (OD) nodes are classified into M macro flows (MF), and the route is changed to each MF. The time is divided into time slots (TS) having a fixed period T, and the MF is designed using the generated traffic amount R _{f, t} of each flow f in each TSt in the evaluation period 1 ≦ t ≦ L. here,

をR_f,tの評価期間における平均値

R _{f, t} during the evaluation period

と定義する。

It is defined as

  Hereinafter, TE control objects such as individual flows and MFs are collectively referred to as “TE control objects”.

Hereinafter, as a first embodiment, the approach to minimize the traffic variation amount V _x of generating MF speed M and MF by greedy algorithm, as a second embodiment, the traffic variation pattern is generated by the cluster analysis We propose two approaches to optimize the extraction ratio when generating MFs by extracting flows from each group of close flows.

[First Embodiment]
In the present embodiment, in the parameter input unit 101 of the flow aggregation apparatus having the configuration shown in FIG. 1, the number of generated MFs M and the amount of traffic generated in each flow f in each time slot (TS) t within the evaluation period are used as parameters. R _{f, t} is input to the MF generator 120.

  Hereinafter, the processing of the MF generation unit 120 will be described.

Considering that one of the MF traffic fluctuation amount V _x and the maximum traffic amount Q is optimized when the number M of generated MFs is given. When present between departure node pair which has flow F present, these combinations of classifying into M MF M ^F number exists, in order to rapidly increase with the increase of F and M, shown in Figure 2 below An approximate optimal solution is derived using a procedure based on greedy arithmetic. However, F _m is defined as a set of flows classified as MFm.

Step 101) Initialize F _m of each m of 1 ≦ m ≦ M to an empty set φ.

Step 102) Select a flow f from the flow set F in descending order of the characteristic value Y _f .

  Step 103) The characteristic value P used as the objective function of the MF set when the selected flow f is classified into each MFm of M is calculated in consideration of only the flows already classified as MF.

Step 104) Classify the flow f into MFm * that minimizes P (F _m = F _m + {f}).

  Step 105) The processing of Steps 102 to 104 is repeated until the entire flow of F is classified as one of the MFs.

The following five (a) to (e) are considered as the characteristic value P used as an optimization objective function and the characteristic value Y _f of f that determines the order of allocation of the flow f to the MF.

(A) Minimization of sum of MF traffic volume dispersion:
For the purpose of reducing the MF traffic fluctuation amount V _x , the total dispersion Φ of the traffic amount during the evaluation period of each MF is set to P. That is,

に設定する。ただしX_m,tは各MFmの各TS tにおける既分類フローの発生トラヒック量の総和であり、

Set to. Where X _{m, t} is the sum of the traffic generated by the classified flows in each TS t of each MFm,

は、以下のX_m,tの評価期間に渡る平均値である。

Is an average value over the following evaluation period of X _{m, t} .

そして、Y_fを以下のように各フローfのトラヒック量の変動係数に設定する。

Then, Y _f is set as a variation coefficient of the traffic amount of each flow _f as follows.

（ｂ） MF最大ピークトラヒック量の最小化：
MF最大トラヒック量Qを低減することを目的に、各MFの評価期間における最大トラヒック量の全MFにわたる最大値ΨをPに設定する。すなわち、

(B) Minimizing MF maximum peak traffic volume:
For the purpose of reducing the MF maximum traffic amount Q, the maximum value Ψ over the entire MF of the maximum traffic amount in the evaluation period of each MF is set to P. That is,

に設定する。そして、以下のように、Y_fを各フローfの最大トラヒック量に設定する。

Set to. Then, Y _f is set to the maximum traffic amount of each flow f as follows.

（ｃ) MF最大総トラヒック量の最小化：
MF最大トラヒック量Qを低減することを目的に、各MFの評価期間における総トラヒック量の全MFにわたる最大値ΩをPに設定する。すなわち、

(C) Minimizing MF maximum total traffic volume:
In order to reduce the MF maximum traffic amount Q, the maximum value Ω over the entire MF of the total traffic amount in the evaluation period of each MF is set to P. That is,

に設定する。そしてY_fを、以下のように、各フローfの総トラヒック量に設定する。

Set to. Then, Y _f is set to the total traffic amount of each flow f as follows.

（ｄ） MF最大ピーク比率の最小化：
経路の安定性を考えたとき、各MFの最大トラヒック量ではなく、全トラヒック量に対して占める割合の最大値を最小化することも考えられる。そこでMF最大トラヒック量Qを低減することを目的に、各MFの各TSにおけるトラヒック量比率の最大値

(D) Minimization of MF maximum peak ratio:
When considering the stability of the route, it is also conceivable to minimize the maximum value of the proportion of the total traffic volume, not the maximum traffic volume of each MF. Therefore, in order to reduce the MF maximum traffic volume Q, the maximum value of the traffic volume ratio in each TS of each MF

をPに設定する。即ち、Fを全フローの集合、

Set to P. That is, F is the set of all flows,

をTStにおける総トラヒック量とすると

Is the total traffic volume in TSt

に設定する。そして、以下のようにY_fを各フローfのピークトラヒック量比率に設定する。

Set to. Then, Y _f is set to the peak traffic amount ratio of each flow f as follows.

（ｅ） MFトラヒック量分散の総和の最小＋ピーク比率制約：
MFトラヒック変動量V_xの低減を目的としつつ、MF最大トラヒック量Qに制約上限を考慮することで、双方のバランスをとる。目的関数として用いる特性値Pは上記の式(1)と同一であるが、図２のステップ１０３，ステップ１０４で各フローfを各MFに収容する際に、ピークトラヒック量比率

(E) Minimum sum of MF traffic volume dispersion + peak ratio constraint:
While aiming to reduce the MF traffic fluctuation amount V _x, the upper limit of the MF maximum traffic amount Q is taken into consideration to balance both. The characteristic value P used as the objective function is the same as the above equation (1). However, when each flow f is accommodated in each MF in step 103 and step 104 in FIG.

が、予め与えた制約上限値を超えないMFのみを候補として考える。Y_fはやはり(ａ)の「MFトラヒック量分散の総和の最小化」と同様に、各フローfのトラヒック量の変動係数に設定する。

However, only MFs that do not exceed the constraint upper limit value given in advance are considered as candidates. Y _f is also set to the variation coefficient of the traffic amount of each flow f in the same manner as the “minimization of the sum of MF traffic amount dispersion” in (a).

[Second Embodiment]
In the present embodiment, an example of generating a macro flow using clustering will be described.

  In the present embodiment, the MF generation unit 102 receives the number of clusters K and the weight parameter β from the parameter input unit 101 in addition to the inputs of the first embodiment.

The both MF traffic variation V _x and MF maximum traffic volume Q strictly perform good Pareto optimal MF aggregation, it is difficult in terms of needs computation time to examine all possible aggregation patterns. However in order to generate a MF with reduced both V _x and Q, by aggregating different flow as far as time fluctuation pattern capable in the same MF, it seems effective to improve the statistical multiplexing effect. From this point of view, this embodiment examines an MF generation method using clustering.

  The MF generation method based on this approach consists of three steps: (I) flow cluster classification, (II) optimal design of the extracted traffic ratio from each flow cluster, and (III) aggregation of flows into the MF based on the optimal extracted traffic ratio. Consists of Each step of FIG. 3 will be described below.

Step 210) Cluster classification of flows:
By clustering each flow f using an L-dimensional vector R _f whose elements are the generated traffic volume R _{f, t} in each TSt, flows with similar traffic time-series patterns in the evaluation period are grouped into the same cluster. Classify. As a cluster classification method, a typical k-means method is used as a non-hierarchical cluster analysis method. The k-means method starts from a state in which the classification target is randomly classified into K clusters given from the parameter input unit 101, and each classification target is assigned to the cluster having the closest distance from the center of gravity of each generated cluster. Repeat the reassignment process until the solution converges.

However, _if the size of R _{f, t} differs greatly from flow to flow, clustering with R _{f, t} as a vector element as it is makes it difficult to classify flows with similar time series patterns into the same cluster. . Therefore, the size of each element of R _f R _f

で除して1に正規化したベクトル

Vector normalized by 1 divided by

に基づき、同様に各クラスタcの重心r_c={r_c,1,r_c,2,…,r_c,L}を正規化した値

Similarly, the normalized value of the centroid r _c = {r _{c, 1} , r _{c, 2} , ..., r _{c, L} } of each cluster c based on

をクラスタcの重心とみなしてクラスタ分類を行う球面k-means法を適用する。以下に、図４と共に、球面k-means法の手順を示す。

Is applied as a spherical k-means method that classifies the cluster as a center of gravity of cluster c. The procedure of the spherical k-means method is shown below with FIG.

  Step 211) For each flow f

を要素にもつL次元正規化ベクトルを算出する。

An L-dimensional normalized vector with element as is calculated.

  Step 212) Center of gravity of each of K clusters c

をL次元空間の正規化ベクトルからランダムに初期設定する。

Is initialized randomly from the normalization vector in L-dimensional space.

  Step 213) Each flow f is

と重心

And center of gravity

との距離

Distance to

が最小となるクラスタcに割り当てる。

Is assigned to the cluster c that minimizes.

  Step 214) Normalized centroid of each cluster c

を再計算する。

Is recalculated.

  Step 215) Repeat Step 213 and Step 214 until the cluster configuration converges.

Step 220) Extracted traffic ratio design from cluster:
A set of flows classified into K clusters c generated in the previous procedure is defined as G _c . Then, consider classifying the traffic of a certain ratio α _{m, c} (0 ≦ α _{m, c} ≦ 1) from each cluster c into M MFm. That is, when the total traffic volume at each TSt of each cluster c is Z _{c, t}

図５に示すようにα_m,cZ_c,tのトラヒックをMF mのTS tに分類する。そのため各MF mのTS tにおけるトラヒック量をM_m,tとすると、

As shown in FIG. 5, the traffic of α _{m, c} Z _c, t is classified into TS t of MF m. Therefore, if the traffic volume at TS t of each MF m is M _{m, t} ,

となる。そして各MFの各TSにおけるトラヒック量の最大値を最小化する以下の線形計画問題を解くことで、分類比率α_m,cを最適設計する。

It becomes. The classification ratio α _{m, c} is optimally designed by solving the following linear programming problem that minimizes the maximum amount of traffic in each TS of each MF.

ステップ２３０）フローのMF集約：
生成した各クラスタを構成するトラヒックを任意の粒度で分割できる場合には、ステップ２２０において、MFの最大トラヒック量を最小化する最適解が得られる。しかし実際にはフローの粒度でのみMF集約が可能であるため、各クラスタを構成するトラヒックを任意の粒度で分割することはできない。そこでフローfをMF mに割当てる場合に1を、そうでない場合に0をとる二値変数をx_f,mとし、各TS tにおいて、x_f,mから定まる各MF mのクラスタcからの抽出トラヒック量

Step 230) MF aggregation of flows:
If the generated traffic constituting each cluster can be divided at an arbitrary granularity, in step 220, an optimal solution for minimizing the maximum traffic volume of the MF is obtained. However, since MF aggregation is actually possible only at the flow granularity, the traffic constituting each cluster cannot be divided at an arbitrary granularity. So, let x _{f, m} be a binary variable that takes 1 if flow f is assigned to MF m, and 0 otherwise. Extraction from each cluster MF m determined from x _{f, m at} each TS t Traffic volume

と、ステップ２２０で導出した分類比率α_m,cから算出されるトラヒック量α_m,cZ_c,tとの差分の全期間にわたる総和sが最小化するようx_f,mを設計する整数計画問題を考える。

And an integer program for designing x _{f, m} so that the sum s over the entire period of the difference between the traffic amount α _{m, c} Z _{c, t} calculated from the classification ratio α _{m, c} derived in step 220 is minimized Think about the problem.

The above MF accommodation policy is expected to reduce the MF traffic fluctuation amount V _x , but at the same time, to reduce the MF maximum traffic amount Q, which is the other design target, Q calculated from x _{f, m} And the following integer programming problem that minimizes the weighted linear sum of s and the maximum value S over all MFs (β is a real parameter in the range 0 ≦ β ≦ 1).

＜生成MFに関する性能評価＞
以下では、米国の実験用NWであるInternet2（登録商標）で取得された4時間にわたるフロー情報の公開データを用いて、第１の実施の形態で述べた貪欲算法に基づく５つのMF生成法と、第２の実施の形態で述べたクラスタリングMF生成法の各々を用いてMFを生成した結果を分析する。

<Performance evaluation on generated MF>
In the following, using the public data of flow information over 4 hours acquired by Internet2 (registered trademark), an experimental NW in the United States, five MF generation methods based on the greedy calculation method described in the first embodiment and The result of generating the MF using each of the clustering MF generation methods described in the second embodiment is analyzed.

・ Evaluation conditions
Internet2 (registered trademark) is composed of 9 nodes. The traffic volume generated in each TS for 5 minutes between 72 OD pairs is set to R _{f, t} , and the first 24 TS is set as the evaluation target period. To do. As an example, only the results between Kansas City-Los Angeles and Chicago-Washington out of 72 OD pairs are shown below. Since the effect of TE cannot be expected for flows with a short length, only flows for which traffic has been observed in three or more TSs during the evaluation period are targeted for MF aggregation. The number of flows used for the evaluation was 1,101 between Kansas City and Los Angeles, and 8,488 between Chicago and Washington.

  In the greedy calculation method in the first embodiment, when the peak ratio constraint is added to the minimization of the sum of the MF traffic volume variance, the value obtained by multiplying the maximum traffic volume in each TS of each flow by 1.2 The upper limit of the allowable traffic volume of MF was set.

  In the method using clustering in the second embodiment, the number of clusters K can be arbitrarily set. However, when verification is performed by setting K to various values, the number of flows included in one cluster is 10. When K below is extremely large, or when K is extremely small such as 2 or 3, MF generation results have been greatly deteriorated, but if it is set to a value in the meantime, change K But there was no big impact on the results. Therefore, for each OD pair, K is 5 if the number of MF aggregation flows in the evaluation period is less than 500, 10 if 500 or less and less than 1,000, 20 if 1,000 or less and less than 5,000, or 5,000 or more. In the case of, each was set to 30. Therefore, K = 20 between Kansas City and Los Angeles, and K = 50 between Chicago and Washington. The solution of the optimization problem in step 220 and step 230 of the MF generation method using clustering was derived using CBC MILP Solver 2.7.5. The weight parameter in the objective function was set to β = 1/3.

-Macro flow generation result:
FIG. 6 plots the average variation coefficient Φ of the traffic volume of the MF generated by each method when the number of MFs M is set to various values ranging from 3 to 50 for Kansas City-Los Angeles. Similarly, FIG. 7 shows the result of Chicago-Washington. However,

である。第１の実施の形態で述べた、目的関数がMFトラヒック量の総和、MF最大ピークトラヒック量、MF最大総トラヒック量、MF最大ピーク比率、そしてMFトラヒック量分散の総和+ピーク比率制約、の5つの貪欲算法による方法を、各々、G/var、G/peak、G/sum、G/pr、G/var+pbと表記する。MFへの集約を行わず全フローを個別に扱った場合(Individualと表記)の結果を点線で示す。

It is. As described in the first embodiment, the objective function is the sum of MF traffic volume, MF maximum peak traffic volume, MF maximum total traffic volume, MF maximum peak ratio, and MF traffic volume dispersion sum + peak ratio constraint. The two greedy arithmetic methods are expressed as G / var, G / peak, G / sum, G / pr, and G / var + pb, respectively. Dotted lines show the results when all flows are handled individually (noted as Individual) without being consolidated into MF.

As M increases, the number of flows that make up each MF decreases, so the statistical multiplexing effect decreases and Φ increases, but Φ is half compared to the case where no aggregation is performed with either method. It can be confirmed that the stability of TE control is improved. In the MF aggregation method, Φ of G / var and G / var + pb that minimize the traffic fluctuation amount of each MF is small, but Φ of G / peak and G / pr that do not consider traffic fluctuation is large. By the way, in all areas of M in Kansas City-Los Angeles and in areas below 15 of M in Chicago-Washington, the Φ of clustering is the largest. Since the aggregation granularity to each MF is a flow, traffic cannot be extracted from each generated cluster with an arbitrary granularity and aggregated to each MF, and step 220 described in the second embodiment. The reason is considered to be an error between the extraction ratio α _{m, c} optimally designed in step 1 and the classification ratio in the MF aggregation actually performed in step 230.

  FIG. 8 shows the cumulative distribution of the variation coefficient of the traffic volume of the MF generated by each method when the number of MFs is set to M = 20. As seen in Figs. 6 and 7, G / var and G / var + pb achieve the smallest Φ, but these methods vary the traffic volume of all generated MFs compared to other methods. It can be confirmed that the coefficient is kept low.

  In FIG. 9, the maximum peak traffic amount Ψ of MF given by Expression (2) is similarly plotted against M for Kansas City-Los Angeles. Similarly, FIG. 10 shows the result of Chicago-Washington. However, the unit of Ψ is the amount of generated traffic (Mbytes) in each TS. Ψ is always non-decreasing by performing aggregation to MF, but Ψ decreases because the number of flows constituting each MF decreases as M increases. After all, Ψ of clustering is large overall. In the region where M is about 15 or more, the difference in Ψ between G / peak, G / pr and Individual is small, and the increase in Ψ when these aggregation methods are used can be ignored. As M increases, the number of flows composing each MF decreases, so Ψ decreases. However, in the region where M is about 20 or more, the ratio of the flow with a large amount of traffic protruding to Ψ increases and M is increased. However, the effect of reducing Ψ is small. Among the MF aggregation methods, G / peak and G / pr that minimize the peak traffic volume have the smallest Ψ in a wide area of M, and G / var + pb also considers the peak ratio as a constraint Shows better characteristics. G / var and G / sum Ψ which do not consider the maximum traffic volume are large.

  Summarizing the above evaluation results, the method using clustering has a smaller effect of suppressing the fluctuation amount and the maximum value of the MF traffic amount as a whole compared with the greedy calculation method. In addition, the method using the class ring requires much longer calculation time than the method using the greedy calculation method. From the above, it is desirable to use a method using a greedy calculation method. Among the greedy arithmetic methods, G / var + pb achieved good evaluation values on both scales. Also, as the number of MFs increases, the amount of change in MF traffic increases, whereas the maximum MF traffic amount converges at about 20, and even if M is increased further, the improvement in results is not expected. Therefore, about 20 MFs are sufficient for the traffic data used for evaluation. For example, when G / var + pb is used at M = 20, the traffic fluctuation amount can be reduced to about 1/5 while suppressing the increase in the maximum traffic amount of TE control to about 15% compared to Individual, and TE Improvement of the link load leveling effect is expected.

<Performance evaluation on TE effect>
By performing TE in units of MF, the number of TE control management targets can be greatly reduced, but on the other hand, the control granularity becomes rough, and there is a concern that the effect of TE may be reduced. Therefore, in the following, the effect of TE achieved in comparison with the conventional case where TE is performed in units of flow and the case where TE is performed in units of MF will be compared and evaluated. The measures to measure the effectiveness of TE are the links over all TSs during the evaluation period, the average traffic volume of each TS (ave.ave), and the average value of the maximum link traffic volume in each TS over the entire period (ave.max). ), The average value of the minimum link traffic volume in each TS over the entire period (ave.min), and the average value of the standard deviation of the traffic volume of each link in each TS over the entire period (ave.sd).

・ Evaluation conditions:
For the TE evaluation, a random NW with 100 nodes, the degree of each node being 2 or more, and an average of 4 was used. And for each of 100 × 99 = 9,900 OD pairs, 4 hours (48TS) of OD pairs selected at random from 72 Internet2 (registered trademark) OD pairs used for the above MF generation evaluation. The traffic amount R _{f, t} of each flow was generated during the simulation period. However, in many flows included in the flow data, traffic occurs only in a very small number of TSs of about 1 to 2 slots, but in order to obtain the effect of TE, only flows with a certain length are targeted. Therefore, only the flow that generated traffic at 4TS or higher was targeted for TE.

Actually, it is necessary to generate MF based on the prediction of future traffic volume, but this paper does not discuss the prediction method, and assumes that the prediction is performed accurately, and traffic of all 48 TSs. The amount was used to generate MF. TE also calculated 30 K-shortest paths (30 paths with the shortest number of hops) for each OD pair at the start of the simulation, and made TE path candidates. In other words, at the start of the simulation, MFs were generated by the methods described in “Performance evaluation on generated MF” above based on the traffic amount R _{f, t} of all 48 TSs. Then, for each generated MF, the path that minimizes the maximum traffic volume of all links on the NWs of all 48 TSs, one by one, in descending order of the maximum traffic in each TS, is selected from the 30 candidate paths. Selected and set. When TE is not used (NonTE), the shortest hop route is always set. In addition, the number of MFs M applied to each OD pair of the evaluation NW was set to 20.

-Link load leveling effect by TE:
To analyze the impact of the introduction of MF on the effectiveness of TE, four measures of link load can be used: when TE is not used (NonTE), when TE is performed for each flow (MicroTE), and Clustering , G / var, G / var + pb Compared with the case where TE is performed with the unit of MF generated by each proposed method. Table 1 summarizes the four evaluation values for each of the five methods (unit: Gbytes / slot). In the case of NonTE, the shortest hop route is always used, so ave.ave is minimized, and increases by about 25% by MicroTE.

しかしMicroTEはNonTEと比較して、ave.maxを30%程度、ave.sdを55%程度低減しており、TEによるリンク負荷平準化効果が確認できる。図１１に各TSにおける最大リンク負荷を、図１２にリンク負荷の標準偏差をプロットするが、評価に用いた全てのTSにおいて、この傾向が確認できる。

However, compared to NonTE, MicroTE has reduced ave.max by about 30% and ave.sd by about 55%, confirming the link load leveling effect of TE. FIG. 11 plots the maximum link load in each TS, and FIG. 12 plots the standard deviation of the link load. This tendency can be confirmed in all TS used for the evaluation.

  The difference in load leveling effect is small among the three MF generation methods. TE by MF increases ava.ave by about 10% compared to MicroTE, but decreases ave.max by about 5% and ave.sd by about 50%. When TE is performed in units of MF, although the control granularity is rough compared to MicroTE that performs TE in individual flow, the amount of traffic fluctuation of TE control target is suppressed as described in “Microflow generation result” above. The effect of being done seems to be a factor.

  As described in “Evaluation conditions” in “Performance evaluation on TE effect” above, each TE control target such as individual flow and MF is considered in descending order of the maximum traffic volume in the evaluation, and the TE control target to which the route is allocated later is considered. Do not select the route. For example, there are two route candidates A and B between the same arrival and departure nodes, and in two TS1 and TS2, flow a is (100, 20) traffic, flow b is (0, 80) traffic, flow, respectively. Consider the case where c generates (80, 50) traffic. When TE is performed for each flow, and when the path a is first assigned to the flow a, the path B is assigned to the flow b to which the path is assigned next. When the remaining flow c is assigned to route A, the maximum link load is 180, and when assigned to route B, the maximum link load is 130. Assigned. On the other hand, for example, when the flow a and the flow b are aggregated into one MF (a + b), if the route A is assigned to the MF (a + b) to which the route is first assigned, the remaining flow c Since the route B is assigned to, the maximum link load in this case is 100.

  As shown in this example, as a result of assigning a route without considering a TE control object to which a route is assigned later, the possibility that the maximum link load increases increases as the traffic fluctuation amount of each TE control object increases. As seen in Figs. 6 and 7, since the amount of traffic fluctuation of MF is smaller than that of individual flows, it seems that the maximum link load and link load distribution are suppressed when MF is targeted for TE control. It is. Therefore, for example, by performing path control in units of MF generated using the method of the present invention with M = 20, the link load leveling effect is improved by about 50% compared to the case of performing individual flows in units. The number of TE control targets can be reduced to about 1/50 to 1/400.

  The processing in the MF generation unit 102 shown in the first and second embodiments is constructed as a program, installed and executed on a computer used as a flow aggregation device, and distributed via a network. Is possible.

  The present invention is not limited to the above-described embodiments, and various modifications and applications are possible within the scope of the claims.

100 Flow Aggregation Device 101 Parameter Input Unit 102 MF Generation Unit 103 MF Output Unit

Claims (7)

A flow that aggregates from OD network address flow to MF unit in a network that can set a route by traffic engineering (TE) in units of macro flow (MF), which is a set of arbitrary traffic flows between pairs of identical arrival and departure (OD) nodes. An aggregation device,
A parameter that inputs the number of generated MFs and a parameter including the amount of traffic R _{f, t} generated for each flow f in each TSt with an evaluation period of 1 ≤ t ≤ L by dividing the time into time slots (TS) with a fixed period T Input means;
MF generation means for classifying F flows existing between each OD node pair into M MFs using the generated traffic amount R _{f, t} of each flow f,
The MF generating means is
Given the number of generated MFs M for the parameter, a set of flows in which F _m is classified as MF m in order to optimize either the MF traffic fluctuation amount V _x or the MF maximum traffic amount Q When defining
1) A first process of initializing F _m of each m of 1 ≦ m ≦ M to an empty set φ;
2) a second step of selecting the flow f from the flow F in descending order of the characteristic value Y _f ;
3) A third process of calculating the characteristic value P used as the objective function of the MF set when the selected flow f is classified into each MFm of the M in consideration of only the flows already classified into the MF;
4) A fourth step of classifying the flow f into MFm * having the minimum characteristic value P (F _m = F _m + {f});
The flow aggregating apparatus that generates M MFs by repeating the processes from the first process to the fourth process until all the flows of F are classified into any of the MFs. The MF generating means is
As an objective function P for optimization and a characteristic value Y _f of f that determines the order of allocation of flow f to MF, X _{m, t} is each of each MFm for the purpose of reducing the MF traffic fluctuation amount V _x The total amount of traffic generated by the classified flows in TS t

Is the average over the evaluation period of X _{m, t}

When

The total variance Φ of the traffic amount during the evaluation period of each MF calculated in the above is P, and Y _f is the coefficient of variation of the traffic amount of each flow f

Means for minimizing the sum of MF traffic volume dispersion, set as
As an objective function P for optimization and a characteristic value Y _f of f that determines the order of allocation of flow f to MF, for the purpose of reducing the MF maximum traffic amount Q,

The maximum value Ψ across all MFs of the maximum traffic volume in the evaluation period of each MF calculated in step S is set to P, and Y _f is set to the maximum traffic volume of each flow f.

MF maximum peak traffic volume minimizing means, set as
For the purpose of reducing the MF maximum traffic amount Q,

The maximum value Ω across all MFs in the evaluation period of each MF calculated in step を is P, and Y _f is the total traffic amount of each flow f.

MF maximum total traffic amount minimizing means, set as
As the objective function P of the optimization and the characteristic value Y _f of f that determines the allocation order of the flow f to the MF, for the purpose of reducing the MF maximum traffic amount Q, F is a set of all flows,

Total traffic volume at TSt

And when

The maximum traffic volume ratio calculated for each MF in each TS

Is set to P and Y _f is the peak traffic volume ratio of each flow f

MF maximum peak ratio minimizing means, set to
When the respective flows f in the third process and the fourth process are accommodated in each MF, the peak traffic amount ratio

However, only the MF that does not exceed the preliminarily given constraint condition value as candidates, minimizing the sum of the MF traffic amount difference and setting the peak ratio constraint means for setting the Y _f as the traffic variation coefficient of each flow f,
The flow aggregation device according to claim 1, comprising any of the following: A flow that aggregates from OD network address flow to MF unit in a network that can set a route by traffic engineering (TE) in units of macro flow (MF), which is a set of arbitrary traffic flows between pairs of identical arrival and departure (OD) nodes. An aggregation device,
Generated MF number M, time is divided into time slots (TS) of a fixed period T, the generated traffic amount R _{f, t} of each flow f in each TSt of the evaluation period 1 ≦ t ≦ L, the number of clusters K, Parameter input means for inputting parameters including weight parameter β,
MF generation means for classifying F flows existing between each OD node pair into M MFs using the generated traffic amount R _{f, t} of each flow f,
The MF generating means is
Cluster classification means for clustering each flow f into K clusters c using an L-dimensional vector R _f having the generated traffic amount R _{f, t} in each TSt as an element;
Extracted traffic ratio design means for obtaining the traffic ratio by minimizing the maximum value of the traffic volume in each TS of each MF from each cluster c,
MF aggregation means for aggregating flows into MF based on the traffic ratio;
A flow aggregation device. The cluster classification means includes:
Using the L-dimensional vector R _f having the generated traffic amount R _{f, t} in each TSt as an element, the flows having similar traffic time series patterns in the evaluation period by the spherical k-means method are the same. Including means for classifying into clusters,
The extracted traffic ratio design means includes:
When a set of flows classified into K clusters c is G _c, and traffic with a certain ratio α _{m, c} (0 ≦ α ≦ 1) is classified into M MF m from each cluster c , The total traffic volume in each TSt of each cluster c

Linear programming problem that minimizes the maximum traffic volume in each TS of each MF

Including means for optimally designing the classification ratio α _{m, c} ,
The MF aggregation means is
A binary variable that takes 1 when allocating the flow f to MF m, and 0 otherwise, is x _{f, m,} and at each TSt, extraction of each MF m determined from x _{f, m} from cluster c Traffic volume

And the maximum value S over the entire MF of the sum s over the entire period of the difference from the traffic amount α _{m, c} Z _{c, t} calculated from the classification ratio α _{m, c} derived by the extracted traffic ratio design means, Integer programming problem to minimize weighted linear sum with Q calculated from x _{f, m}

The flow aggregation apparatus according to claim 3, further comprising means for designing x _{f, m} by solving A device that aggregates OD network address flows in MF units in a network that can be routed by traffic engineering (TE) in units of macroflows (MF), which are a set of arbitrary traffic flows between pairs of identical departure and arrival (OD) nodes. The flow aggregation method in
Parameter input step for dividing the number of generated MFs and time into time slots (TS) with a fixed period T and inputting parameters including the amount of traffic R _{f, t} generated for each flow f in each TSt in the evaluation period 1 ≦ t ≦ L When,
The computer executes an MF generation step of classifying F flows existing between each OD node pair into M MFs using the generated traffic amount R _{f, t} of each flow f,
The MF generation step includes
Given the number of generated MFs M for the parameter, a set of flows in which F _m is classified as MF m in order to optimize either the MF traffic fluctuation amount V _x or the MF maximum traffic amount Q When defining
1) A first process of initializing F _m of each m of 1 ≦ m ≦ M to an empty set φ;
2) a second step of selecting the flow f from the flow F in descending order of the characteristic value Y _f ;
3) A third process of calculating the characteristic value P used as the objective function of the MF set when the selected flow f is classified into each MFm of the M in consideration of only the flows already classified into the MF;
4) A fourth step of classifying the flow f into MFm * having the minimum characteristic value P (F _m = F _m + {f});
Summarizing and generating M MFs by repeating the processes from the first process to the fourth process until the entire flow of F is classified as one of the MFs . A device that aggregates OD network address flows in MF units in a network that can be routed by traffic engineering (TE) in units of macroflows (MF), which are a set of arbitrary traffic flows between pairs of identical departure and arrival (OD) nodes. The flow aggregation method in
Parameter input step for dividing the number of generated MFs and time into time slots (TS) with a fixed period T and inputting parameters including the amount of traffic R _{f, t} generated for each flow f in each TSt in the evaluation period 1 ≦ t ≦ L When,
The computer executes an MF generation step of classifying F flows existing between each OD node pair into M MFs using the generated traffic amount R _{f, t} of each flow f,
The MF generation step includes
Clustering each flow f into K clusters c using an L-dimensional vector R _f whose elements are generated traffic R _{f, t} in each TSt;
Obtaining a traffic ratio by minimizing the maximum amount of traffic in each TS of each MF from each cluster c;
Aggregating flows into MF based on the traffic ratio;
The flow aggregation method characterized by including. Computer
The flow aggregation program for functioning as each means of the flow aggregation apparatus of any one of Claims 1 thru | or 4.

Images