JP2021005387A5 - - Google Patents

Description

Today, enterprises with large computer infrastructure can easily accommodate the maximum number of concurrent users they expect to use their applications on any given day. Infrastructure requirements tend to be very large. This is seen as a means of ensuring that users do not suffer performance degradation or unreasonable downtime. This approach may seem reasonable, but unfortunately it has led to a surge in servers that are mostly unused or underutilized for long periods of time. This overprovisioning ultimately means that companies will incur unnecessarily increased capital investment and operating costs throughout the year in terms of licensing, power, heat, cooling, and data center space, to name just a few. Means to.

Several different methods of time series forecasting have been proposed so far to anticipate the needs of enterprise computer infrastructure. This previous study includes using classical models such as linear regression, exponential smoothing, and autoregressive integrated moving averages (ARIMA), support vector machines (SVM), artificial neural networks (ANN), and fuzzy logic. It extends to highly computational intelligence non-linear methods. The k-nearest neighbor (k-NN) algorithm has also recently been extended to time series conjecture, but existing studies in this area focus on individual historical data without reference to any additional relevant metrics. ing. Therefore, most applications use only one-dimensional load curve shapes to identify similar nearest neighbors in the past, while being constrained to one-dimensional univariate time series, between large dimensional feature spaces. Correlation is often ignored. However, many of the measured computer hardware metrics are closely related to each other, so knowledge of history extracted from one dimension is compared to that collected from time-series groups that share similar behavior. Then it is inappropriate.

In one general aspect, the invention is a network resource and / or infrastructure need for a corporate computer system that employs a network server to host the resources (applications, data, etc.) required by network users. Regarding computer implementation systems and methods for anticipating. Based on expectations, network resources can be scaled or provisioned accordingly. That is, for example, the state of the network server can be dynamically adjusted to meet the needs of the user while reducing excess capacity. The predictive technique of the present invention is also applicable to cloud computing environments. Based on expectations, the cloud server pool is dynamically scaled, so the scale of the system meets changing demands and avoids wasting resources when the system is under load.

Forecasts are generated based on historical data about user requests for network resources. In one scenario, multivariate k-nearest neighbor (k-NN) predictions are used based on the grouping of request variants grouped based on those metrics according to correlation analysis, so between the metrics within the same group. Dependency is high. A multivariate k-NN algorithm is then executed for each group to generate a multi-step forward prediction. This technique reduces the impact of other irrelevant metrics so that the prediction performance improves with the prediction speed. Another prediction technique uses prediction and stripe sampling techniques. By predicting the number of future requirements in a unit time, analyzing the characteristics of the requirements, classifying the requirements accordingly, and determining the slice sampling strategy, detailed future workload data can be generated. Yet another approach is an adaptive scheduling algorithm based on demand pattern prediction. This algorithm determines the adaptability of the system to different workloads and adjusts the cloud resources used by the system based on past and present system loads to reduce waste while meeting the dynamic load of the system. To be able to.

These and other advantages from embodiments of the invention will become apparent from the description below.

Various embodiments of the invention are described herein as examples in connection with the accompanying drawings.

FIG. 3 is a block diagram of a corporate computer system according to various embodiments of the present invention.

FIG. 5 is a flow chart of a process flow for determining a server number recommendation for the enterprise computer system of FIG. 1 according to various embodiments of the present invention.

It is a graph which illustrates an example of the prediction method for the univariate k-NN method by various embodiments of this invention.

FIG. 3 is a block diagram of the predictive computer system of FIG. 1 according to various embodiments of the present invention.

FIG. 3 is a flow chart of a process flow executed by the computer system of FIG. 3 according to various embodiments of the present invention.

An example of a controller of the resource prediction computer system of FIG. 1 according to various embodiments of the present invention is illustrated.

An example of a typical load (eg, number of requests) on a network server of a corporate computer system over a day is shown.

FIG. 6 is a diagram of a transfer function for the controller of FIG. 6 according to various embodiments of the present invention.

In one general aspect, embodiments of the invention can be used to predict future computer resource needs for an enterprise. One exemplary enterprise computer system 10 is illustrated in FIG. The corporate computer system 10 illustrated in FIG. 1 includes several local area networks (LANs) 12 interconnected with a wide area network (WAN) 14. Each LAN 12 can include some client computers 16 and some network servers 18. The network server 18 may host computer resources such as computer programs, data, storage devices, and printers for the client computer 16 within or from another LAN 12, for example, depending on the implementation mode.

The resource prediction computer system 20 executes a company's resource prediction based on multivariate time series (MTS) data for the network server 18 according to various embodiments, and the MTS data is stored in the database computer system 22. The resource prediction computer system 20 and the MTS database system 22 are shown in FIG. 1 as being connected to the WAN 14 for illustration purposes, but one or both of them may be in one of the illustrated LANs 12. Can be included. They can also connect to different LANs 12 and WAN 14s within the corporate network. The resource prediction computer system 20 is one or more interconnected computer devices such as servers (s), mainframes (s), workstations (s), and / or any other suitable computer device. It may be implemented as. Each such computer device of the resource prediction computer system 20 may include one or more processors 24 and one or more memory units 26. The memory unit 26 may include both primary computer storage (eg RAM and ROM) and secondary computer storage (eg HDD, SSD, flash). The processor (s) 24 is a computer instruction (s) stored in the memory unit 26, such as the variable grouping module 30 and the k nearest neighbor (k-NN) search module 32, as shown in FIG. 1 and further described below. For example, it may include a microprocessor (s) running software). For purposes of illustration, the illustrated resource prediction computer system 20 includes only one computer, and only one processor 24 and one memory unit 26, but the invention is not limited thereto. It should be understood that the resource prediction computer system 20 can be scaled as needed.

The MTS database 22 stores time-series computer utilization and hardware statistics for the network server 18 of the enterprise system 10. Statistics can include the values of several variables associated with each user resource request (such as:), as follows:
-User name of the user requesting the resource-Request start time-Request end time-Request total time-Request operation time-Requested process or resource-ID of the network server that responded to the request
-Geographical location of the network server that responded to the request-CPU utilization-Network server primary memory (eg RAM) utilization-Total of network server read / write operations to disk IO (disk memory or other secondary computer memory) )

The MTS database 22 runs database management system (DBMS) software and optionally uses one or more database servers, including a suitable RAID disk array and / or any other suitable data storage mechanism. It may be implemented. Usage and hardware statistics may be sent to the MTS database 22 using CSV files and / or any other suitable data format from the network server 18. Each network server 18 may send its periodic statistics to the MTS database 22 and / or one or more network servers 18 in the LAN 12 collect statistics about a plurality of network servers 18 in the LAN 12. The collection can be sent to the MTS database 22. The MTS database 22 may be implemented as, for example, an Oracle or SQL database, or any other suitable database.

In various embodiments, in order to predict the future computer resource needs of an enterprise, the resource prediction computer system 20 is dependent on variables such that each related variable is a member of one cluster and has the same cluster. Group related variables into a cluster so that they are more likely but very low compared to the variables in other clusters. The processor 24 may perform this function by executing the software of the variable grouping module 30. After the cluster is determined, the resource prediction computer system 20 uses k-NN search in various embodiments to calculate the prediction. The processor 24 may perform this function by executing the software of the k-NN search module 32.

FIG. 2 is a process flow of a prediction process that may be performed by the resource prediction computer system 20 along with the MTS database 22 according to various embodiments. In step 50, the MTS database 22 receives and stores time series usage and hardware statistical variables for the enterprise network server 18. Data is collected for periodic time increments, such as every 5 minutes, every 10 minutes, or any other suitable time increment. Then in step 52, the resource prediction computer system 20 preprocesses the MTS data stored in the database 22 to remove or reduce noise and / or use conventional data preprocessing techniques to preprocess the data. Can be scaled to standard normal distribution data.

When dealing with high-dimensional time series, the overall k-NN estimation based on all dimensions can have large errors due to interference between irrelevant metrics. However, some of the measured computer hardware metrics are closely related to each other and share similar behavior, so if the prediction is simply extracted from the univariate time series, the relevant information is lost. It will be. In order to utilize the knowledge aggregated from all related variables, before performing the k-nn search, the higher dimensional MTS is decomposed into smaller groups, where the dependencies between variables within the same group are high. However, the dependency with variables in other groups is very low. This operation is performed by the processor 24 of the resource prediction computer system 20 that creates a variable cluster by running the software of the variable grouping module 30 in step 54 of FIG. Such decomposition also significantly narrows the model space, thus accelerating subsequent k-NN searches. In various implementations, variable grouping involves two steps: (i) correlation matrix construction, and (ii) variable clustering.

式中、ｘ_ｉ、ｙ_ｉは、未処理の変数を指し、バー付きのｘ、ｙは、変数の平均を指す。 Given d variables and an MTS of length T, the correlation coefficient can be used to measure the association between the two variables. In one embodiment, Spearman's rank correlation coefficient, which is a monotonic nonparametric measure of the relationship between two variables, is used. Spearman's correlation does not assume that both datasets are normally distributed. For a sample of size n, the unprocessed variables X _i and Y _i are converted to ranks x _i and y _i , and Spearman's rank correlation coefficient ρ is calculated from these as follows.

In the equation, x _i and y _i refer to unprocessed variables, and x and y with bars refer to the average of the variables.

すべての変数の中で、対での相関係数は、ｎ×ｎ相関行列を形成するために計算される。以下の表１は、特定のネットワークサーバホストを用いた実験で構築された例示的な相関行列である。これは、より濃い灰色を有する対称行列が強い相関を示すことに留意されたい。

Keep in mind that one variable can affect other variables after a time lag, so for each pair of variables i, j, the algorithm can search from 0 to a predefined "MaxLag" period. The maximum coefficient of ρ _ij can be selected as follows.

Among all variables, the paired correlation coefficient is calculated to form an n × n correlation matrix. Table 1 below is an exemplary correlation matrix constructed in an experiment with a particular network server host. Note that this is strongly correlated with symmetric matrices with a darker gray color.

Based on the correlation matrix constructed above, a partitioning method can be applied to decompose n variables into clusters. In various embodiments, an Affinity Propagation (AP) clustering algorithm is used in this step. AP is a clustering algorithm based on the concept of "message passing" between data points. Unlike other known splitting methods such as k-means and k-means, APs do not need to determine or estimate the number of clusters before running the algorithm. Therefore, the grouping result may change as the input hardware metrics change, making it more suitable for this embodiment. Further details about the AP are incorporated herein by reference in their entirety, Brendan J. et al. Free et al. , "Clustering by passing messages beween data points," Science 31: 972-976 (2007).

となり、式中、ｓ（ｉ，ｋ）は、サンプルｉとｋとの間の類似度である。アルゴリズムは、収束するまで２つの行列の計算を実行する。表２は、表１の行列に基づくクラスタリング結果を示す。

この例で分かるように、各変数０～６は、グループ／クラスタのうちの１つに属する。 The AP algorithm creates a cluster by sending a message between sample pairs until it converges. There are two categories of messages sent between points. One is the responsibility that sample k should be a sample of sample i, represented by r (i, k). The other is the availability that sample i should be selected as a sample of sample k, represented by a (i, k):

In the equation, s (i, k) is the degree of similarity between samples i and k. The algorithm performs the calculation of the two matrices until it converges. Table 2 shows the clustering results based on the matrix of Table 1.

As can be seen in this example, each variable 0-6 belongs to one of the groups / clusters.

After the variable cluster has been determined, in step 56 the k-NN algorithm is used to make the desired predictions. For purposes of illustration, the k-NN search algorithm is initially illustrated only for univariate situations and for values one step ahead based on a constant number of previous data measured at constant width time intervals. Will be done. In the case of t = (1, 2, ..., T), the predicted range h = (1, 2, ..., H) is predicted from the point T x _{T + h} (where x is ⌒). Consider a finite and equidistant time series x _t . First, a set of feature vectors is created to represent the features of the recent state of the time series. A segment of equal length m is a sequence of m consecutive observations x ^m _t (where x is with →) = [x _t , x _t-1 , ..., x _{t- (m-1).} ] ^T is considered as a vector sequence x ^m _t (where x is with →), where m is a predetermined integer called the sampling length. Note that the vector overlaps with all t = (m, m + 1, m + 2, ..., Tm), the so-called m history.

Then, in various embodiments, a given distance metric, such as the Euclidean distance metric, is used to identify similar patterns of behavior in the past, using all m histories and the last observed vector _x ^mt . (Here, x is with →) = [x _T , x _T-1 , ..., x _{T- (m-1)} ] The distances to ^T are calculated respectively.

式中、ｎｅｉｇｈ_ｊ，ｈは、近傍のｊの範囲ｈに続く値であり、ｗ_ｊは、重みを表す。図３は、一変量ｋ－ＮＮ手法のための予測方法を図示するグラフである。基準期間（３４１～３８１日目）に最も近い２つの近傍（１７１～２２１日目と１～５１日目）を示す。 Finally, the paired distances to all m histories are ranked and subsequent observations of the k vectors with the minimum distance to the target feature vector are selected respectively. Each of the k nearest neighbors can be locally weighted and then aggregated to form the prediction x _{T + h} (where x is ⌒).

In the equation, negh _{j and h} are values following the range h of nearby j, and w _j represents a weight. FIG. 3 is a graph illustrating a prediction method for the univariate k-NN method. The two neighborhoods (days 171 to 221 and days 1 to 51) closest to the reference period (days 341 to 381) are shown.

であり、式中ｘ^ｍ _Ｔ，ｉ（ここでｘは→付き）は、次元ｉで時間Ｔにおいて最後に観測されたベクトルを表す。同様に、距離メトリクスは、距離メトリクスの合計として定義することができる（例えば、すべてのｄ次元におけるユークリッド距離

ｍ履歴にわたって探索することによって、ＭＴＳの現在の（または他の参照）状態に近いグローバルなｋ個の最近傍を見つけることができ、それらの各々は、ｍ個の連続した観測のｄ個のベクトルによって構成される。ｋ個の最近傍の範囲ｈを有する後続の値は、参照ｋ×ｄ行列として表すことができる。

各次元に対する予測値は、対応する次元での参照項目を集計することによって生成される。

ここで、各次元の近傍は、結果を形成するためにそれぞれ重み付けされる。一実施形態では、重みは、勾配降下アルゴリズムによって得られ、新しいデータと一致するように徐々に調整される。予測は、時間範囲の累積ユーザリソース要求とそれらの要求に対処するために必要なサーバリソースを予想することができる。したがって、予測は、時間範囲ステップに対して予測されたサーバワークロードをもたらす。拡張されたワークロード予想を得るために、このプロセスを追加の時間範囲ステップに対して繰り返すことができる。 For multivariates, the reference (eg, up-to-date) feature vector and m-history for univariates are extended in multiple dimensions. Let X ^d be a d-dimensional multivariate time series (MTS). To predict the future of MTS at time T, the target feature matrix is

In the equation, ^xm _{T, i} (where x is with →) represents the last observed vector in time T in dimension i. Similarly, distance metrics can be defined as the sum of distance metrics (eg, Euclidean distance in all d dimensions).

By searching over the m history, we can find the global k nearest neighbors that are close to the current (or other reference) state of the MTS, each of which is a d vector of m consecutive observations. Consists of. Subsequent values with k nearest neighbor ranges h can be represented as a reference k × d matrix.

Predicted values for each dimension are generated by aggregating the reference items in the corresponding dimension.

Here, the neighborhood of each dimension is each weighted to form a result. In one embodiment, the weights are obtained by a gradient descent algorithm and are gradually adjusted to match the new data. Prediction can predict cumulative user resource requests over a time range and the server resources needed to handle those requests. Therefore, the prediction yields the predicted server workload for the time range step. This process can be repeated for additional time range steps to obtain enhanced workload expectations.

In step 58 of FIG. 2, the resource prediction computer system 20 can recommend the number of servers in a future time range step. Such server number predictions can be used for a number of useful purposes. One such purpose is to allow an enterprise to provision its servers 18 based on an expected number of servers. Several calculated metrics, including the following metrics, may be used in such capacity planning.
-CPU / memory / disk overload factor and underload factor. Calculations of these metrics can be performed by averaging CPU, memory, and disk overload and underload values (above or below the respective overload and underload thresholds, respectively). The threshold value can be, for example, 90% in the case of overload and 10% in the case of underload.
-CPU / memory / disk overload time rate and underload time rate. These metrics are the length of time that the CPU / memory / disk is overloaded (above the overload threshold) or underloaded (below the underload threshold) compared to the overall period. It can be calculated based on the ratio of the values.
-Overall overload time rate and underload time rate. In this calculation, if any of the CPU / memory / disk metrics for the server is overloaded or underloaded, then the server is considered to be overloaded or underloaded in some cases.
These metrics can be predicted by the resource prediction computer system 20 for future time steps based on the k-NN predicted CPU / memory / disk load described above (see step 56).

These expected overload / underload metrics are input parameters for various recommended methodologies for recommending hardware resources for enterprise computer systems, such as the recommended number of servers 18 (see Figure 1). Can be used. Three such recommended methodologies are the so-called "capacity-based server number recommendation" (C-SNR), "usage-based server number recommendation" (U-SNR), and "adaptive server number recommendation" (A-). SNR). In C-SNR, the number of requests that can be processed by the server is predetermined. Therefore, once the total number of requests is predicted, the number of servers can be easily derived. In the U-SNR, the most likely time for a user to execute a server request is known based on the user's usage history and classification. Since the average load value of the user for the request is also recognized based on the classification of the user, the load expected to be brought to the system by the user's request can be predicted. The number of servers can be determined by applying an allocation algorithm that seeks to maximize server capacity. The A-SNR uses the ratio (actually the threshold) to determine whether to add a server to send a new user session and a new server. When a new request is made, the first attempt is for the currently running server to handle the request. Requests are assigned to the current server based on the potential load level estimated from the classification results. If a new request cannot be assigned to the currently active server due to an overload condition, a new server is added to handle the request. Similarly, once the request is complete, the remaining workload can be rebalanced across the servers 18.

Regardless of which of these metrics is used, or any other metric, the Resource Prediction Computer System 20 is an appropriate number for the enterprise for a particular time period based on the expected user workload. You can anticipate the server 18. The resource prediction computer system 20 can communicate these server number recommendations to the network server 40, which acts as a broker (see FIG. 1) for the network server 18. Resource Prediction Based on the decisions communicated from the computer system 20, the broker server (s) 40 may start several servers 18 in the LAN 12 at various points in time (fully active, operating mode) and / or low power. It is possible to determine whether to enter a mode (eg, idle mode or sleep mode) and instruct the server 18 accordingly. In this way, some of the network servers 18 can be put into low power mode during periods when they are needed and not expected, based on expected usage and / or load patterns. For example, with reference to FIG. 1, the resource prediction computer system 20 can determine the expected number of network servers 18 required for a particular time period. The conjecture can be for a network server 18 within one LAN 12 or a network server 18 across multiple LANs 12. The resource prediction computer system 20 can send this number to the broker server 40, which is a high power mode capable of operating at the appropriate time to the LAN 12 and / or various servers 18 in the other LAN 12. It can be instructed to be in (can handle the user's resource request) or in low power mode (cannot handle the user's resource request). The network server 18 can assume the power mode instructed accordingly. In this way, a certain number of network servers 18 can be put into low power mode in order to save energy and associated costs when the expected network resource needs are low. Conversely, when the expected network resource needs are high, a sufficient number of network servers 18 will be prepared to handle the expected user-requested load.

In various embodiments, the processor 24 (when running the k-NN search module software 32) selects the k and m parameters for the k-NN search prior to the k-NN search. Module 32 may use a high climbing algorithm to adjust the k and m parameters to minimize prediction errors for different data sets. An iterative algorithm of a high climbing algorithm starting with an arbitrary solution to the problem (in this case, the k and m parameters) then seeks to find a better solution by incrementally modifying a single element of the solution. do. If the change produces a better solution, then incremental changes to the new solution are made that are repeated until no further improvement can be found. In other embodiments, the k and m parameters are reestablished.

The relationship between predictions with K-nearest neighbors can change over time. Therefore, the weights are preferably adjusted over time. It is assumed that Train _t = [x _t-1 , x _t-2 , ..., X _tn ] represents a learning set of size n at time t. x _t is the true value at time t and y _t is the predicted value for time t. Based on the previously obtained weights, y _t can be predicted based on Train _t . After knowing the true value of x _t , the weight can be updated based on Train _{t + 1} = [x _t , x _t-1 , ..., x _{tn + 1} ], and this new weight is y _{t + 1} Used for prediction of.

Another common problem is the dynamic change in computing resources required by an enterprise. This is a big waste if the enterprise maintains the number of servers for the highest workload, as the workload does not need such a large number of servers during low hours. In addition to maintaining enough servers to meet user demands, workloads need to be predicted to achieve the goal of saving as much energy as possible. Therefore, various embodiments of the present invention provide workload generation methods based on prediction and stripe sampling techniques. Future detailed workload data can be generated by predicting the number of future requirements in a unit time, analyzing the characteristics of the requirements, classifying the requirements accordingly, and determining the slice sampling strategy. Future workload data can provide a basis for dynamically adjusting server resources.

In one general aspect, the prediction and stripe sampling techniques can be divided into steps. First, a time series prediction model is applied to predict the number of requests in a unit time. Based on the k-nearest neighbor (KNN) algorithm, K subsequences similar to the current subsequence in the requested number time series data are searched. A linear regression model can then be applied to synthesize the predictions given by the K nearest neighbor sequences into the final prediction result.

Second, we analyze the characteristics of the requirements. Each request consumes different resources (CPU load, memory, etc.). In addition, it is necessary to predict the length of each request. Analysis can be divided into three types according to various embodiments.
a. Correlation analysis. Correlation analysis can be performed on attributes with different requirements. When two attributes are correlated, they depend on each other, while two uncorrelated attributes do not affect each other.
b. Classification request. By classifying requests by relevant attributes, you can classify requests into different types, such as short-term requirements with low internal storage costs, long-term requirements with low internal storage costs, and long-term requirements with high internal storage costs. Can be divided into.
c. Periodic analysis. The negative effects of periodic factors should be eliminated. The requirements show a significant difference between working and non-working hours. If it is desired to sample the request, a request within a similar periodic time should be selected for sampling.

Third, using a combination of slice-based sampling and period-based sampling, requirements are selected from historical data according to the proportion of each category, and these request data are combined into a final forecast.

This process has the advantage that the workload is measured based on multiple metrics, thereby overcoming the limitations of a single metric. Another potential advantage is to make more accurate predictions by modeling the workload brought about by each requirement, taking into account the number of future requirements and the impact of each requirement on various metrics. As such, this method overcomes the limitations of the overall workload prediction method.

FIG. 4 is a diagram of a resource prediction computer system 20, and FIG. 5 is a resource prediction to implement this prediction and stripe sampling strategy to generate future workload traces according to various embodiments of the present invention. It is a figure of the process flow executed by the computer system 20. As shown in FIG. 3, according to such an embodiment, the resource prediction computer system 20 may include a prediction and stripe sampling module 60 for generating future workload traces. The prediction and stripe sampling module 60 may be implemented as software code stored in the memory (s) 26 and executed by the processor (s) 24 of the resource prediction computer system 20. The exemplary process flow of FIG. 5 can be performed by the resource prediction computer system 20 when the code of the prediction and stripe sampling module 60 is executed. Further, the database 22 can store data regarding past requests by users of the network server.

要求数が現時点のそれらと類似している場合を見つけるために、最新のｍ個の観測値からなるサブシーケンスを描画することができる。

もう一方の観測は、次のように表すことができる。

In step 62, the number of requests per unit time is predicted. The problem of predicting how many requests will be sent by the user per unit time can be modeled as a time series prediction problem. Therefore, using K nearest neighbor classification algorithms, it is possible to find a subsequence similar to the current subsequence from the past data stored in the database 22. The predictions obtained based on these similar subsequences can then be combined together to produce the final prediction result. Assuming that the time interval is the unit time T, the number of requests generated in each unit time can be expressed as N = [n ₁ , n ₂ , ..., n _p ], and in the equation, n ₁ Is the number of requests generated within the i-th unit time, and p can be expressed as follows.

In order to find out when the number of requests is similar to those at the present time, a subsequence consisting of the latest m observations can be drawn.

The other observation can be expressed as follows.

A plurality of subsequences _{Ni, i + m-1} = [ni _i , ..., _{ni + m-1} ] can be found from the past data most similar to _Now . Evaluation criteria can be used to measure the similarity between past subsequences and current subsequences. Let Dis (X, Y) be a distance function for measuring the similarity between two time series. The distance function Euclidean distance can be used, which is as follows.

The distance between each past subsequence and the current subsequence can be calculated. Past sequences can be divided into subsequence sets {N _{1,1 + m-1} , N _{2,2 + m-1} , ..., N _{p-2 * m + 1, pm} }. The distance can then be calculated to find i in the range 0 <i≤p-2 * m + 1, which ensures that the value of Dis (N _{i, i + m-1} , _Now ) is minimized. The subsequence _{Ni, i + m-1} most similar to the current one can then be determined.

For all past subsequences in the set, the distance di = Dis (N _i , _{i + m-1} , _Now ) is calculated and the distance set D = {d ₁ , d ₂ , ..., d. _{p-2 * m + 1} } can be obtained. After all the elements in D have been sorted, the closest distance [d _a , db, ..., d _k ] can be determined, and the distance [a, _b , ..., D k] can be determined. The index of [・, k] is the index of the most similar subsequence.

The subsequences immediately following these K nearest neighbor sequences are preferably combined to generate a final prediction. It is preferable to consider two factors. First, similarity should be considered in the model, and in general, the higher the similarity, the greater the weight. Second, subsequence appearance times, generally more recent appearance times, should have greater weight. Therefore, in various embodiments, the k subsequences are sorted according to their similarity to the current subsequence and the time of occurrence. You can then apply a linear regression technique to learn the weights for the kth prediction.

Next, in step 64, the required feature or attribute is analyzed. Attributes can include request data such as request duration, disk IO, CPU and / or memory usage, as described above. This step 64 can include three sub-steps according to various embodiments, a correction analysis in step 66, a requirement classification in step 68, and a periodic analysis in step 70. In correction analysis step 66, each requirement generally results in different workloads for different computer resources. Therefore, multiple features should be used to characterize the workload. One question is whether these workload features are correlated. For example, you should know if the claim that as CPU consumption decreases, so does disk input / output consumption applies to an application.

この相関係数の絶対値が大きいほど、相関度は高い。値が０に近い場合、２つの特徴は、ほぼ独立していることを示す。 According to various embodiments, the Pearson correlation coefficient can be used in this step. Assuming that the request workload has two specific features, this is two time series X = [x ₁ , x ₂ , ..., X _N ] and Y = [y ₁ , y ₂ , ,. ..., y _N ] can be further expressed, and in the equation, the value of the i-th request can be expressed as (x _i , y _i ), and the correlation between these two features is as follows. Can be expressed in.

The larger the absolute value of this correlation coefficient, the higher the degree of correlation. When the value is close to 0, the two features are almost independent.

Regarding the requirement classification step 68, by observing the distribution of the feature values with respect to the number of features, some features may have a plurality of peaks, which indicates that the feature values can be classified. Requests can be categorized into different types so that they can be sampled from the past according to the proportion of the class. During the classification process, classification should be made for correlated features. In this step, a bisected K-means clustering algorithm based on the classical k-means clustering algorithm can be used to avoid falling into a locally optimal solution. This clustering algorithm can be divided into three steps. The first step is to find the clustering center, the second step is to calculate the distance between each point and the cluster center and put each point in the nearest cluster, the third step. Is to calculate the average of all cluster coordinates as the new cluster center. The bisected K-means clustering method first puts all the points into one cluster, then divides one cluster into two, and finally two clusters that can reduce the maximum error. It can be further enhanced by choosing to divide into. This process can be repeated until the number of clusters is equal to the number of Ks given by the user.

Cyclic factors are important factors in forecasting. Since the requests to the server are made by people, it is clear that the number of requests during and after working hours is different. Therefore, the effects of cyclical factors should be considered in periodic analysis step 70. In one embodiment, the Fast Fourier Transform (FFT) can be used to calculate the cycle length of the required feature sequence. The average value of the characteristic x _i at the time interval t can be obtained. Next, at a sampling rate f _s = 1 / t, the time series f (n) and n = 0, 1, 2, ..., N-1 in the equation can be determined.

を得ることができ、式中、

ｋ＝０，１，２，・・・，Ｎ－１である。ＤＦＴを得た後、各ｋは、次のように表すことができる、離散周波数値に対応する。

シーケンスが周期的である場合、フーリエ変換関数のスペクトルは、周期の逆数でヒットし、衝撃値によってｆ（ｎ）の周期を得ることができる。単一の特徴によって表される要求負荷については、対応するサイクル時点から要求をサンプリングすることだけが必要とされる。複数のカテゴリへと分類されている要求については、あらゆる時間単位における各カテゴリの比率を予測することができ、これはステップ６２で導入されたＫ－ＮＮ手法を採用することによって行うことができる。 Assuming that the discrete Fourier transform (DFT) of f (n) is F (k),

Can be obtained during the ceremony,

k = 0, 1, 2, ..., N-1. After obtaining the DFT, each k corresponds to a discrete frequency value, which can be expressed as:

When the sequence is periodic, the spectrum of the Fourier transform function hits with the reciprocal of the period, and the impact value can give the period of f (n). For demand loads represented by a single feature, it is only necessary to sample the demand from the corresponding cycle point in time. For requests classified into multiple categories, the ratio of each category in any time unit can be predicted, which can be done by adopting the K-NN method introduced in step 62.

From the results of the analysis, in step 72, the requirement data can be extracted from the past data in order to simulate future requirements. It is assumed that periodicity analysis has shown that the cycle length is _Tperid . In the classification, the requests are divided into K classes, and the past request collection N _history is divided into several subsets.

It is assumed that p requests will occur in the future at a certain time interval _tpredict , and the predicted ratio of the k-st class is wk. Therefore, the number of requests for category k is pk = wk * p. Random numbers can be generated to select a request within a size range of the past subset N _{tredict, k ,} pk. Requests of all categories can then be integrated into a collection of simulated future requirements. The time of occurrence of each past request is not fixed within each unit time interval. The relative start time can be expressed as:
t _unit = t _actual -kT, k = 0,1,2, ...
During the simulation of future request sequences, the actual request time can be expressed as _tpredict = t _unit + _tperiod .
By the above method, future workload traces can be generated. Also, based on future workload traces, server count recommendations can be made for future time periods to meet the needs of enterprise users based on expected workloads (see step 58 in FIG. 2). ).

One advantage of the above solution is that the workload is measured based on multiple metrics, which overcomes the limitation of a single metric. Another advantage is to make more accurate predictions by modeling the workload that each requirement brings, taking into account the number of future requirements and the impact of each requirement on different metrics. Therefore, this solution overcomes the limitations experienced with the overall workload prediction approach.

Another potential useful use of the expected usage pattern is for enterprises migrating to virtual desktops hosted by cloud computing vendors that offer desktops as a service (DaaS). Enterprises can perform forecasts to determine expected DaaS resource sizes such as CPU count, RAM capacity, and / or storage capacity. As such, an entity has too much cloud computing resources for its needs (and thereby overpaying for those needs), and too few loud computing resources. It is possible to avoid having only possession (and thereby not having the required resources for that user).

Another useful use is to determine the number of network servers 18 that an enterprise needs for a particular workload or resource. Based on the company's expected usage pattern for a particular resource that indicates the number of servers required (according to one of the recommended methods above), the company can procure the appropriate number of servers. For example, when the system is under heavy load, the controller adds a server to the system, for example, a virtual server, to share the request. When the system load is low, the controller returns extra resources to reduce costs. FIG. 6 illustrates a controller 120 (eg, the resource prediction computer system 20 shown in FIG. 1) for making such a decision.

The requirements of the system can be described as time series data, i.e., the number of requirements changes cyclically over time. For example, daytime demands are generally higher than nighttime demands. In various embodiments of the invention, the controller 120 anticipates future demands and adjusts the system according to different policies. The general steps of the process are described below in connection with FIG.

The k-NN prediction unit 122 uses the k-NN algorithm to predict future demands. In various embodiments, the predictor 122 only predicts the request pattern during the next period, not a specific number of periods. Request patterns can be divided into three types: increase, decrease, and variation. The request pattern is classified as an increase if the number of requests increases over the X1 period (eg, 3 periods) or if it increases by more than Y1% (eg, 40%) during the Z1 period (eg, Z1 = 2). can do. The next pattern after the increase pattern will also be classified as an increase pattern until they meet a continuous non-increase period. During the Z2 period (eg, Z2 = Z1 = 2), the number of requests continues to decrease over the X2 period (eg, X2 = X1 = 3), or decreases by more than Y2% (eg, Y2 = Y1 = 40%). If so, the request pattern is classified as a decrease. Other patterns are classified as variability, and in various embodiments, the period following the increase or decrease pattern can also be classified as increase or decrease, in some cases until a decrease or increase period occurs. ..

FIG. 7 shows a typical service system load for a day. In this example, the sample period is every 15 minutes. In this example, the request data is on the increase from 7:00 am to 8:00 am. The number of requests increases rapidly with a slight decrease period within this time frame. From 8:00 am to 9:00 am, the number of requests is in a decreasing pattern, and the number of requests decreases for most of this one-hour period. From 10 am to 2 pm (ie, 14:00 on the time / time axis), the request data is in a variation pattern as the number of requests continues to fluctuate during this 4 hour period. Other periods can be categorized in the same way. After dividing the request pattern into three types, the k-NN prediction unit 122 predicts the request pattern for the next control cycle. The number of requests for all control cycles is recorded. The prediction unit 122 groups several consecutive points (for example, five consecutive points) into one group, and uses it as sample data for classifying the latest data group. Next, the k-NN prediction unit 122 searches for a similar group from past data. In various embodiments, the request exhibits normal periodicity, so the k-NN predictor 122 searches for similar periodic times in the history. The variance between the two groups is used to determine the similarity between them. The k-NN prediction unit 122 then selects K most similar groups (eg, K = 5). After finding these similar groups, the k-NN prediction unit 122 obtains the pattern for the next period of the group and takes the most frequent pattern as a prediction of the request pattern for the next period.

ここで、ｕ（ｔ）は、時刻ｔにおける制御出力であり、これは追加するサーバ数であり、ｅ（ｔ）は、システムの負荷と理想的な負荷との偏差であり、Ｋは、比例係数であり、ｒ´（ｔ）は、時間ｔにおける予測要求数であり、Ｔ_ｃは、制御周期であり、Ａは、予測影響係数である（予測が制御プロセスにどの程度影響するかを決定する）。式の右側の第１の部分（すなわち、Ｋｅ（ｔ））は、従来の比例制御である。第２の部分（すなわち、ＡＫ（ｒ´（ｔ＋Ｔ_ｃ）－ｒ´（ｔ）））は予測の影響を制御プロセスに導入し、予測影響係数によって出力に影響を与える。 The three patterns have different characteristics. As a result, the decision-making unit 126 can apply different control policies. In the increasing pattern, it takes time to get the server resource from the virtual provider and wait for the service to start, so it is necessary to add the server in advance. The following formula can be used to calculate the required number of additional servers.

Here, u (t) is the control output at time t, this is the number of servers to be added, e (t) is the deviation between the system load and the ideal load, and K is proportional. It is a coefficient, r'(t) is the number of prediction requests at time t, T _c is the control period, and A is the prediction influence coefficient (determines how much the prediction affects the control process). do). The first part on the right side of the equation (ie, Ke (t)) is conventional proportional control. The second part (ie, AK (r'(t + T _c ) -r' (t))) introduces the influence of the prediction into the control process and influences the output by the prediction influence factor.

In the reduction pattern, server reduction rarely results in delays, so there is little or no need to reduce servers in advance. Therefore, the decision-making unit 126 can close the server and return it to the server when the load on the system is low. The calculation can be performed according to the equation u (t) = Ke (t). Compared to the above equation for the increasing pattern, the decreasing pattern equation contains only the proportional control part.

ｅ（ｔ）≧０のとき、意志決定部１２６は、比例制御によってサーバを追加するだけである。ｅ（ｔ）＜０の場合、これはサーバ数を低減する必要があり得ることを意味し、意思決定部１２６は、最後の３つの制御期間（または他のいくつかの制御期間）にわたる負荷レベルを収集し、これらの期間すべての間システムが低負荷状態にあった場合、コントローラは、（３つの制御期間のうちの）負荷が最も大きい制御期間を選択して、減少パターンについて上述したようにクローズして戻すサーバ数を計算する。制御ポリシーの目的は、システムの安定性を確保しながらサーバ数を安定させることである。 In the variation pattern, the decision-making unit 126 must carefully deal with changes in the load of the system because the requirements may vary frequently. If the decision-making unit 126 only changes the server according to the load, the system is unstable because the server is paid on an hourly basis in addition to the energy consumption cost generated by repeatedly starting and stopping the server. This can lead to a great deal of waste. Therefore, for the fluctuation pattern, the decision-making unit 126 can use the following equation.

When e (t) ≧ 0, the decision-making unit 126 only adds a server by proportional control. If e (t) <0, this means that the number of servers may need to be reduced, and the decision-making unit 126 has a load level over the last three control periods (or some other control period). If the system was in a low load state for all of these periods, the controller would select the control period with the highest load (out of the three control periods) and as described above for the reduction pattern. Calculate the number of servers to close and return. The purpose of the control policy is to stabilize the number of servers while ensuring the stability of the system.

Since adding a server may have some delay, the controller 120 may include a Smith predictor 124 to compensate for the effects of the delay. The Smith predictor 124 can receive as input the number of servers available / purchased but not yet used. Based on this input, Smith Predictor 124 estimates the capacity of these servers and compensates for these reserved capacities against the current load on the system. The compensated system load is then delivered to decision-making unit 126 for use in the decision-making process. The principle of Smith Prediction Unit 124 is that, in various embodiments, a compensation portion is introduced into the feedback of the system, which pre-populates the decision-making unit 126 with the delayed feedback amount.

特性方程式は、次のとおりである。

特性方程式内の遅延項は、スミス予測部の補償後に除去され、これはシステムに対する遅延の影響を低減することが分かる。 FIG. 8 is a transfer function of the controller 120 according to various embodiments of the present invention. In this figure, G ₀ (s) e- ^ts represents a delay in the system, and the transfer function of the Smith predictor is G ₀ (s) (1-e- ^ts ). The Smith prediction unit 124 is connected to the k-NN prediction unit 122 G _τ (s) to form a controller with pure time delay compensation. With the addition of the Smith predictor, the system's closed-loop transfer function looks like this:

The characteristic equation is as follows.

It can be seen that the delay term in the characteristic equation is removed after the Smith predictor compensates, which reduces the effect of the delay on the system.

式中、Ｔは、サンプリング期間の長さであり、ｓは、この期間の平均サーバ数であり、ｒは、この期間の完了した要求数であり、ｓＴ／ｒは、単位時間あたりの単位サーバの効率を表し、Ｌは、修正係数で、調整率を決定する。Ｌの値が大きいと、Ｋの応答はより速くなるが、変動にもつながる場合がある。 The learning unit 128 collects system load and state information and uses that information to adjust the proportionality factor K online in the decision unit 126. The proportionality factor K determines how the controller 120 reacts to the load of the system. The higher the K value, the more adjustments are made, which makes the system more volatile. The learning unit 128 can adjust K online according to the formula.

In the equation, T is the length of the sampling period, s is the average number of servers in this period, r is the number of completed requests in this period, and sT / r is the unit server per unit time. Represents the efficiency of, and L is a correction coefficient, which determines the adjustment rate. The larger the value of L, the faster the response of K, but it can also lead to fluctuations.

The controller 120 can perform the above-mentioned processes using appropriate software instructions stored in the member 26 and is executed by the processor (s) 24 of the computer system 20 (see FIG. 1). For example, the memory 26 may include software modules for k-NN prediction, Smith prediction, learning, and decision making, which are executed by the processor (s) 24 to perform the process described above. The result can be provided to the broker computer system 40, which in some cases can ramp up or down the network server. By incorporating these four parts (122, 124, 126, 128) and applying different control policies for different request patterns, embodiments of the invention are suitable for scalable cloud systems for most of the time. It can be controlled, guarantees the performance of the system, and reduces unnecessary costs.

Yet another useful use is the simulation of virtual user workloads. Historical and future time data based on real client workloads can be generated for simulated load testing of virtual systems.

Accordingly, in various embodiments, the present invention is directed to computer systems and related computer implementation methods that anticipate the needs of network resources for enterprise computer systems. According to various embodiments, the computer database system receives and stores multivariate time series (MTS) performance data for a plurality of network servers. The MTS performance data includes data for a plurality of d performance variables for a plurality of network servers for a series of past sampling times. A programmed computer system (eg, resource prediction computer system 20) groups variables in MTS performance data into two or more variable groups such that each of the performance variables in MTS performance data belongs to a variable group. .. The programmed computer system then computes the forecasts of future workloads on the network servers of the corporate computer system by computing the forecasts of the variables in one or more future time range steps. The programmed computer system calculates the prediction by performing a step that includes:
(I) Finding k nearest neighbors to the reference state of the MTS performance data using the k nearest neighbor search algorithm applied to two or more variable groups, and (ii) weighting k nearest neighbors. Calculate the average. The programmed computer system is then put into operation mode to address resource demands by users of the enterprise computer system at each of one or more future time range steps, based on calculated predictions. You can determine the recommended number of network servers required for this.

In various implementations, performance variables for multiple network servers include at least variables that indicate CPU load per unit period, primary computer memory utilization, and secondary computer storage input / output (IO) operations. Also, the steps for grouping variables are the step of calculating the correlation matrix showing the correlation between each pair of performance variables by the programmed computer system and the step of using the clustering algorithm by the programmed computer system. It may also include a step of determining a variable group based on. The correlation matrix may include a Spearman correlation matrix, and the clustering algorithm may include an affinity propagation algorithm.

Further, in the step of finding the k nearest neighbors closest to the reference state of the MTS performance data, (i) the distance between each of the vector representing the reference state of the MTS and each of the plurality of vectors representing the MTS data is previously set. It may include calculating in sampling time and (ii) determining k vectors for the past sampling time with the shortest distance to the vector representing the reference state of MTS. The calculated distance may be the Euclidean distance.

In addition, the programmed computer system is the recommended number of network servers required for the enterprise to be in operating mode in order to address resource demands by users of the enterprise computer system for at least one of the future time range steps. Data indicating the above may be transmitted. Therefore, one or more broker computer systems may instruct network servers so that the recommended number of network servers is in an operating mode for addressing resource requests by users for at least one future time range step. ..

In another general aspect, the invention relates to computer systems and related computer implementation methods for predicting future workloads of network servers over future time periods. A programmed computer system (eg, a resource prediction computer system 20) predicts the number of requests p for a future period based on a sort of k most recent subsequences of the period, and the enterprise per unit time T. The number of requests from a computer system user to a network server is most similar to the current subsequence for a recent period. The programmed computer system then classifies past requests (based on the data stored in database 22) into two or more request type classes based on the attributes of the request. The programmed computer system then predicts the ratio of requirements in the future period for each of the two or more request type classes, based on the ratio of past requirements in each of the two or more request type classes. The programmed computer system then determines the periodicity for one or more request attributes for the request type class. The programmed computer system will then be configured so that the p samples have a predicted ratio of each of the two or more request type classes, and the p samples are based on the periodicity of the request type class in the future. Sample p past requests so that they are from the same request cycle point as the period of. Finally, the programmed computer system synthesizes p sampled past requests to obtain a workload trace for a network server in a future period.

According to various implementations, one or more broker systems can adjust the state of the network server for future time periods based on the predicted future workload. The programmed computer system also performs a correlation analysis of the attributes of past requests, including classifying past requests into two or more classes based on the correlation analysis, based on the attributes of the past . Requests can also be classified into two or more request type classes. The programmed computer system calculates the Pearson correlation coefficient between the pairs of request attributes when performing a correlation analysis and clusters to divide the request into two or more requests based on the request attributes of the request. Algorithms can be used. It is also possible to use the Fast Fourier Transform for periodic analysis to calculate the cycle length of one or more required attributes of two or more classes.

According to another general aspect, the invention relates to computer systems and related computer implementation methods that anticipate the need for network resources for enterprise computer systems. The programmed computer system (eg, the resource prediction computer system 20) determines the predicted user resource request pattern for future time range steps based on the user request data of the enterprise computer system user. The programmed computer system then calculates the recommended number of servers required by the enterprise for future time range steps based on the classification of predicted user resource request patterns for future time range steps. In doing so, the programmed computer system uses the first control policy to calculate the recommended number of servers when the predicted user resource request pattern for future time range steps is classified as increasing. The second control policy is used to calculate the recommended number of servers when the predicted user resource request pattern for future time range steps is classified as diminished.

In various implementations, the first control policy uses the first equation containing the proportionality constant K to calculate the recommended number of servers, and the second control policy uses the second equation containing the proportionality constant K. Use to calculate the recommended number of servers. In that case, the programmed computer system can also adjust the proportionality constant K online for the first and second control policies, based on the efficiency of the unit network server within the unit period. The programmed computer system can also use a third control policy to calculate the recommended number of servers when the predicted user resource request pattern for future time range steps is not classified as increasing or decreasing. .. The third control policy can also calculate the recommended number of servers using the (third) equation containing the proportionality constant K.

In various implementations, the predicted user resource request patterns for future time range steps are categorized as increasing when one or more conditions are met, which one or more conditions are the duration of the predicted number of requests. From the first condition that the number of prediction requests exceeds the first threshold number (N1) of the range step and the first threshold percentage exceeds the second threshold number (N2) of the continuous time range step. Includes the second condition that there is a large increase. Also, one or more conditions for the classification of increase include a third condition that (i) the immediately preceding time range step is classified as an increase and (ii) the future time range step is not classified as a decrease. be able to. Similarly, a predicted user resource request pattern for a future time range step can be classified as a decrease when one or more conditions are met, one or more of which is a contiguous number of predicted requests. The first condition that the number of prediction requests decreased by exceeding the third threshold number (N3) of the time range step, and the second threshold value exceeded the fourth threshold number (N4) of the continuous time range step. Includes a second condition that it is reduced by more than a percentage. There may also be a third condition for the classification of declines, where (i) the previous time range step is classified as decreasing and (ii) the future time range step is not classified as increasing. The programmed computer system also compensates for the capacity of available network servers that are not currently in use when calculating the recommended number of servers required by the enterprise for an increasing future time range step. You can also.

It should be noted, for example, that FIG. 1 illustrates only the components of a corporate computer system 10 sufficient to understand aspects of the invention. The corporate computer system 10 can include a large number of interconnected LANs, WANs, MANs, etc. with wired (eg, Ethernet over twisted pair cable) or wireless (eg, Wi-Fi) communication links. Should be recognized. The network server 18 can handle a large number of different types of resources according to the demands of the enterprise, and the client computer 16 can be any suitable type such as laptops, personal computers, tablet computers, smartphones and the like. It can be a network-enabled end-user computer device. The resource prediction computer system 20 can be implemented by one or more network computer devices. When the resource prediction computer system 20 includes a plurality of computer devices, they can be interconnected by one or more LANs, WANs, MANs, and the like. In addition, the corporate computer system 10 may include an additional broker computer 40 for provisioning the server 18.

The software modules described herein are pythons such that when the processor (s) 24 execute the software program of the module, the processor (s) 24 may perform the functions of the module described herein. It can be implemented in one or more computer programs written in any suitable computer language such as (Python). Suitable processors 24 for executing instructional programs include, for example, both general and special purpose microprocessors. In addition, any computer component described herein can include a single processor or multiple processors. The processor 24 receives instructions and data from read-only memory and / or random access memory.

The present specification contains details of many specific implementations, which should not be construed as a limitation of any scope of the invention or claims, but rather to a particular implementation of a particular invention. It should be interpreted as an explanation of features that may be unique. Certain features described herein in the context of different implementations can also be implemented in combination in a single implementation. Conversely, the various features described in the context of a single implementation can also be implemented in multiple implementations separately or in any suitable secondary combination. Further, the features may be described above as acting in a combination, and may be initially claimed as such, but in some cases one or more features from the claimed combination. It can be cut out, and the claimed combination may be subject to a secondary combination or a modification of the secondary combination.

Similarly, the actions are depicted in the drawing in a particular order, but this is to be performed in a particular order or sequence in which such actions are shown, or all, in order to achieve the desired result. It should not be understood that the illustrated actions need to be performed. In certain situations, multiple task processing and parallel processing may be advantageous. Moreover, the separation of the various system components in the implementations described above should not be understood as requiring such separation in all implementations, and the program components and systems described are generally single. It should be understood that it can be integrated in a product or packaged into multiple products.

In this way, specific implementation embodiments of the subject have been described. Other implementations are within the scope of the following claims. In some cases, the operations described in the claims can be performed in a different order and still achieve the desired result. Moreover, the process shown in the accompanying drawings does not necessarily require the particular order or sequential order shown to achieve the desired result. In certain implementations, multiple task processing and parallel processing may be advantageous.

Claims (12)

A computer implementation method for predicting future workloads of a network server of a corporate computer system over a future period, wherein the corporate computer system hosts multiple networks hosting computer resources for users of the corporate computer system. The above method is equipped with a server.
The programmed computer system predicts the number of requests during the future period based on the sort of k past subsequences, each of which is a unit. It contains a time series of requests per hour, the k past subsequences are the k past subsequences closest to the current subsequence, and the current subsequence is the current time. Predicting and predicting, which is a time series of requests per unit time
The programmed computer system classifies past requests in the k past subsequences into two or more request type classes based on the attributes of the past requests .
The future for each of the two or more request type classes based on the ratio of the past requests of the k past subsequences in each of the two or more request type classes by the programmed computer system. Predicting the ratio of requests within the period of
The programmed computer system determines the periodicity of the past request attributes for the two or more request type classes.
Sampling p past requests by a programmed computer system, the sampled p past requests having a predicted ratio for each of the two or more request type classes. And the sampled p past requests are from the same cycle point in time as the future period, based on the periodicity of the attributes of the past requests in the two or more request type classes. Sampling and
To obtain a workload trace for the plurality of network servers for the future period by synthesizing the p past requests sampled by the programmed computer system, in the workload trace. Includes, captures, and methods of future workloads of said plurality of network servers in said future time period . The method of claim 1, further comprising adjusting the state of the plurality of network servers in the future period based on the predicted future workload. It is possible to classify the past request into the two or more request type classes based on the attributes of the past request.
Performing a correlation analysis of the attributes of the past request by the programmed computer system,
The method of claim 1, comprising classifying the past requirements into the two or more classes by the programmed computer system based on the correlation analysis. The method of claim 3, wherein performing the correlation analysis comprises calculating a Pearson correlation coefficient between the pairs of attributes of the past request. Classification of the past request uses a clustering algorithm for the programmed computer system to divide the past request into the two or more request type classes based on the attributes of the past request. The method according to claim 4, comprising the above. Determining the periodicity for the past request attributes of the two or more request type classes is a fast Fourier transform to calculate the cycle length of the past request attributes of the two or more request type classes. The method of claim 1, comprising using a transform. A system for predicting future workloads of network servers in a corporate computer system over a future period, wherein the corporate computer system hosts multiple network servers that host computer resources for users of the corporate computer system. In preparation, the system
A computer database system that communicates with the plurality of network servers to store data about past requests made by the user to the plurality of network servers.
A predictive computer system that communicates with the computer database system, wherein the predictive computer system
Predicting the number of requests for future time periods based on the sort of k past subsequences, each of the k past subsequences containing a time series of requests per unit time. , The k past subsequences are the k past subsequences closest to the current subsequence, and the current subsequence is a time series of requests per unit time in the current time. Predicting and
To classify past requests in the k past subsequences into two or more request type classes based on the attributes of the past requests .
Percentage of requirements in the future period for each of the two or more request type classes, based on the ratio of the past requirements of the k past subsequences in each of the two or more request type classes. To predict and
Determining the periodicity of the attributes of the past request for the two or more request type classes.
By sampling p past requests, the sampled p past requests have the predicted proportions for each of the two or more request type classes and are sampled . The p past requests made are sampled from the same cycle point in time as the future period, based on the periodicity of the attributes of the past requests in the two or more request type classes. To do and
Combining the p past requests sampled to obtain a workload trace for the plurality of network servers for the future period, the workload trace is for the future period. A predictive computer system, which is programmed to predict the future workload of the network server by performing steps including acquiring and including the future workload of the plurality of network servers. The system. It further comprises a broker computer system that communicates with the predictive computer system, for the broker computer system to adjust the state of the network server in the future period based on the predicted future workload. , The system according to claim 7. The predictive computer system
Performing a correlation analysis of the attributes of the past request,
Based on the correlation analysis, the past requirements are classified into the two or more classes, and the past requirements are classified based on the attributes of the past requirements by performing a step including. The system according to claim 7, which is programmed to classify into two or more request type classes. 9. The predictive computer system is programmed to perform the correlation analysis by performing a step comprising calculating a Pearson correlation coefficient between the pairs of attributes of the past request. The system described in. The past by performing steps that include the predictive computer system using a clustering algorithm to divide the past request into the two or more request type classes based on the attributes of the past request. 10. The system of claim 10, which is programmed to classify the requirements of. The two or more by performing a step in which the predictive computer system includes using a Fast Fourier Transform to calculate the cycle length of the attributes of the past request of the two or more request type classes. 7. The system of claim 7, programmed to determine the periodicity of the request type class with respect to the attributes of the past request.