Today, to create modern systems designed to process Big Data in real time, it is necessary to develop scalable solutions based on the adaptive computing infrastructure, capable of providing prompt and e cient service of incoming requests, as well as storing information obtained from many sources. A signi cant di erence in the paradigm of working with Big data is the lack of a clear data storage structure, as well as the heterogeneity of information, ows coming from di erent sources with high intensity and in large volumes. The scienti c problem, the solution of which is the present study, is to increase the e ciency of routing using modern virtualization technologies and machine learning methods for managing information ows. Solving this problem will allow implementing a data analysis system that can adapt to changes in information ows of high intensity due to self-organization of computing resources, self-learning, algorithmic solutions based on arti cial intelligence.
In recent years, the amount of data that needs to be processed and analyzed grows exponentially. Modern technologies used for data processing require the storage of Big Data. For example, for the application of methods of machine learning or neural networks, it is necessary to ensure the storage of reference samples. Thus, a lot of data that were previously considered useless, not stored for long and not processed in automated order now have a high value. Searching for hidden patterns, trends, emissions, automatic analysis of correlations with data from completely di erent formats and origin helps solve many problems in more simple and quick ways. But the amount of data supplied for analysis, even by a small set of sensors or a server, can be calculated in terabytes per day. Most companies do not have the resources either for real-time analysis or for their long-term storage. One of the possible solutions to overcome this problem is the use of cloud computing technologies.
As a rule, cloud systems are located on the basis of data centers. Such data
centers have signi cant amounts of computing, network and software resources,
allowing to scale cloud solutions for processing Big data [
Basically, large providers use ready-made solutions for deploying a software and hardware platform for cloud computing. The most popular are the systems for managing cloud resources OpenNebula, OpenStack, CloudStack, Doku, etc. However, existing solutions for the deployment of cloud systems are universal. This does not allow to take into account a number of features of working with large data. The key disadvantages of cloud systems when working with Big Data are:
- the lack of the possibility of parallelizing the operation of the network (alternative routing and switching of communication channels); - a large amount of consumed memory and calculations for one analysis cycle.
These shortcomings do not allow e cient processing of Big data by specialized services deployed on the basis of an arbitrarily chosen cloud computing infrastructure.
To solve these problems, providers provide users with ready-made sets of
cloud services in BigData-as-a-Service format. As a rule, BigData-as-a-Service
includes a Hadoop / Spark cluster with storage on HDFS / Hive / Gluster, as
well as NoSQL-services like Redis / MongoDB. However, as the practice of
using such services shows, end-users use the service not only to calculate statistics
or retrieve data. BigData-as-a-Service is often used to build models aimed at
predicting data changes or their automatic classi cation. Such algorithms are
di cult to implement on the Hadoop platform. This is due to the fact that the
execution time of the model in Hadoop can be comparable with the usual
processing on one computing node. Processing data in the cycles Map and Reduce
each time initiates access to the le system. Spark also uses a full mesh network
architecture and pre-build RDD tables that require additional samples [
On the other hand, even large servers with more than 60 cores and terabytes of RAM cannot accommodate the entire amount of data required for processing. For these purposes, as a rule, cluster storage technologies are used. However, this approach is not always applicable to existing methods of analysis and processing of large data (for example, for neural network applications). Distributed execution of neural networks involves great di culties in synchronizing the layers when the layer is divided into sections. Intermediate storage should have a special logic of work, high performance, as well as a small delay in terms of communication over the network.
To solve the identi ed problems, this study proposes the use of approaches that optimize the network infrastructure for the intensive tra c ows that arise during the analysis and processing of Big data. The solution of this task is based on the principles of self-management and virtualization of the resources of the heterogeneous infrastructure of the data center. The proposed solution is based on the hybrid application of software-con gurable components for adaptive routing of data streams, using the functionality of two breakthrough technologies software-de ned network (SDN) and virtual network functions (VNF).
Today, SDN is the most popular and e ective approach to the
organization of the network for the provision of services on the basis of data centers.
The use of this technology is due to a number of advantages. First of all, SDN
greatly simpli es the design and operation of the network, since it allows
centralized intelligent control at the controller level. Secondly, SDN allows network
administrators to quickly con gure and optimize network resources based on an
aggregated set of data collected in a single location. Thirdly, the use of SDN
allows providing protection by dynamically analyzing data ows circulating in a
virtual data center [
Another technology which used to organize a network based on virtual data
centers is the technology of network function virtualization (NFV). Technology
NFV o ers a new way of designing, deploying network services based on the
cloud systems. Virtualization of network functions allows separating network
functions, such as NAT, Routers, and Switches from the hardware level [
The problem of organizing adaptive routing for Big data ows can be solved by formalizing an optimization task to maximize the satisfaction of QoS requirements. The scienti c novelty of the task is that at the moment there are no e ective algorithmic solutions for automatically providing adaptive routing within the allocated network resources for processing Big data. 2
As part of this study, an approach combining adaptive routing in processing Big data ows and providing QoS (including minimum guaranteed throughput and maximum guaranteed delay) is proposed.
Let G (t) = (V; E), be an oriented multigraph describing the current topology of the virtual network at some time t. The set of its vertices V is the union of the set of nodes representing the set V = (vN ode; N F V; vStg), where vN ode virtual computing nodes that process Big data, N F V - virtual network devices (switches, routers, etc.), as well as vStg - virtual data stores in which the data for analysis are directly placed.
Each arc e 2 E, corresponds to some network connection between the vertices Start (e) 2 V and Stop (e) 2 V . There can be several parallel arcs between the two vertices, for example, parallel connections between routers. This provides a variety of alternative routes for data transmission and QoS parameters.
On the set of arcs E two functions are given: 1.Lspeed : E ! R+ [ f0g - a mapping characterizing the current capacity of each arc at time t.
2.Ldelay : E ! R+ [ f0g - The maximum guaranteed delay.
Then each service working in the cloud system and dealing with large data processing has its own mechanism for organizing the communication interaction of tra c ows, which can be represented as the next oriented graph G0 = (V 0; L) (1) where V 0 - a set of virtual nodes on which the components of a service that processes Big data are deployed, L - arcs pointing to data ows between virtual computing nodes, network devices, and a storage system.
To ensure the quality of service in the cloud network, the tasks on the set of arcs L are three functions:
1.lsmpiened : L ! R+ [ f0g - The minimum guaranteed throughput of data ows corresponding to this arc.
2.ldmealaxy : L ! R+ [ f0g - The maximum guaranteed delay.
3.ldaevlgay : L ! R+ [ f0g - estimate the average delay that occurs when processing data ow packets on ports of network devices.
Then the QoS routing and provisioning algorithm must construct a function : L ! P ath (G (t)) such that each data ow l 2 L, is associated with its transmission path r leading from the vertex Start (l) to Stop (l). Here, as P ath (G (t)), the set of routes between any pairs of vertices in the graph of topology G (t), which is actual at time t, is indicated. The function can be described as a vector R = r1; : : : ; rjLj , where ri = (li) is the route for the data ows.
Also, we introduce an additional notation, let : E ! 2l be a function that maps each network connection e 2 E to a set of data ows l 2 L for which the corresponding routes pass through e, that is, (e) = fl 2 Ljr = &e 2 rg.
Obviously, the vector R must contain routes that satisfy the following QoS constraints:
1. The bandwidth of each route ri, taking into account the e ect of other data ows, should not be less than the guaranteed throughput for the ows 8 > > li : 8ri me2irni X <Lspeed e2ri >>:
X lsmpiened (l) l2 fl(ieg) > > ; lsmpiened (li)
2. The total delay of each route ri, taking into account the in uence of other data ows, should not be greater than the guaranteed delay for the ow li : 8ri 8 > X ><ldaevlgay (e) + 2. ri must start at the vertex corresponding to the beginning of arc li and end in the vertex corresponding to the end of li, i.e.:
8ri; Start (ei;1) = Start (li) &Stop (ei;ni ) = Stop (li) : 3. ri must not pass several times through the same vertex, i.e.:
8ri8j; k = 1; nj ; j 6= k ) Start (ei;j ) 6= Stop (ei;k) :
Thus, for e ective organization of adaptive tra c routing when processing Big data ows, it is necessary to solve two tasks:
- automate the construction of rules for routing tra c, based on data based on data on the current state of the network and the services that are running and the processes running in them;
- optimize the rules for routing tra c to ensure the quality of service in the virtual network.
To solve the rst task, it is necessary to perform classi cations of network tra c passing through the corporate network. To solve the second problem, we use the iterative method, in which we distinguish two main stages. At the rst stage, we will perform the clustering of the rules obtained as a result of solving the rst task, and the subsequent deducing of the rules from the clusters. At the second stage, to solve the second problem, we apply the algorithm of con ict-free optimization of the list of routing rules. 3
To implement the presented plan, it is necessary rst of all to determine the model of routing rules used to provide the functioning of a virtual network of a cloud system built for processing Big data.
A routing rule is usually a string consisting of certain characteristics of a network connection. Within the framework of the study, the following characteristics were chosen from a variety of characteristics describing network connections: the IP address of the destination network, the network mask, the IP address of the next router, the output interface, and the metric characterizing the priority of using the route. When selecting these characteristics, the routing rule will generally look like: < dst ip; net mask; gateway; interf ace; metric > (2) where: dst ip - IP address of the destination network; net mask - network mask; gateway - IP address of the next router; interf ace - output interface; metric - the metric that characterizes the priority of using a route in the routing table.
To analyze the demand for communication channels and improve routing e ciency, services that collect statistical information are deployed on virtual network nodes. The list of records about the packets passing through virtual network nodes is represented by the form of the set of the following form:
R = frkg ; k = 1; n where n - is the length of the list of records stored on the virtual network node that provides the operation of the cloud platform.
Let's de ne areas of admissible values for elements of the given vector, namely characteristics of tra c on which collection of statistics of use of communication channels in a network of the software-de ned infrastructure will be carried out. They are represented as a list of lines of the form (2). Each such line is represented as a vector of the form:
rk = frk1; rk2; rk3; rk4; rk5g ; where: k - number in the list of tra c rules; rk1 -{ IP address of the destination network;rk2 -{ network mask; rk3 -{ IP address of the next router; rk4 - output interface; rk5 -{ the metric that characterizes the priority of using a route in the routing table. (3) (4) 3.1
Algorithm for Deriving Routing Adaptive Rules from Clusters After reducing the input data to the required form, we formulate the statement of the problem. Given a lot of records about the tra c passing through the virtual network in the form of a set of the type (4), it is required to build a set of non-con icting rules R = frig ; i = 1; m; m ! min.
This work assumes the solution of the problem by the iterative method in two stages. The rst step is the construction of the initial list of rules R1 by classifying the set X into two classes. The second step consists of making of the set Ropt by clustering the set R1 with the subsequent deduction of the rules from the clusters, that is, the construction of the set R2 with the subsequent optimization of this set by the algorithm of con ict-free optimization, that is, the construction of the set Ropt.
There are many di erent methods for solving this problem. In the framework of this study, we will use the classi cation based on the neural network. This is due to the fact that classi cation is a classic task for neural network methods. The most applicable for solving such a problem is a multi-layered perceptron type architecture. The number of neurons in the input layer is calculated depending on the input data. In this case, the size of the input layer of the neural network will be 99 neurons. It takes 32 bits to write an IP address, and 16 bits to write ports, the number of protocols considered is 7, therefore, 3 bits are required to write them. To train the selected neural network model, we will use the algorithm for backpropagation of the error. 3.2
One of the problems with processing Big data is the constant access to a speci c resource for obtaining data for analysis. As a result of such activity, there is a rapid growth of tra c and as a result of the load on the processor and the increase in the amount of memory consumed on the virtual network nodes involved in organizing the route. This leads to a drop in performance of the cloud network as a whole. In order to reduce computing, it is necessary to solve the task of compiling a list of rules without losing the characteristics responsible for the performance of the network and the cloud system itself. For these purposes, it makes sense to break routing rules into clusters in order to derive new, generalized rules from them. Let's formulate the mathematical formulation of the rules of rewall clustering tasks.
Let there be given a set of rules R = frig; i = 1; m. It is required to compose a sample partition into disjoint subsets called clusters in such a way that each subset consists of objects that are close in some metric. It is necessary to compose a clustering function f (r) : R ! Y , that assigns to each element of the set R an element of the set Y = Y1; Y2; : : : Ym - the set of cluster numbers.
An important aspect in solving the clustering problem is the choice of the distance function, or metric. The metric is a measure of proximity, which is algorithms. As part of the study, the Euclidean distance was taken as the metric by the following characteristics: the source and destination IP addresses, the corresponding ports, the protocol by which the connection is made and the decision on the admissibility of the connection. Thus, the formula for the distance function has the following form:
D (r1; r2) = q a (r1;2 + r2;2)2 + b (r1;3 + r2;3)2; (5)
This function was used with empirically selected parameters a = 0; 55, b = 0; 55.
The next step is to remove the rules from the clustered list of rules. Input data will be a list of rules, broken into clusters Rkl. The output of the algorithm is expected to list the generalized rules Ropt. In the framework of this study, an algorithm was developed. The list of clusters is denoted by u. Then for all clusters from u, we have:
Input: Clustered rule list Rk;l, list of clusters u.
Output: Generalized list of rules Ropt. 1 ri;1 = min (rk;1) =mask = 32 log2 (max (rk;1) 2 ri;2 = min (rk;2) =mask = 32 log2 (max (rk;2) 3 ri;3 = fr1;3; : : : ; rk;3g; 4 ri;4 = fr1;4; : : : ; rk;4g; 5 ri;5 = fr1;5; : : : ; rk;5g. min (rk;1)); min (rk;2));
Algorithm 1: Clustering of adaptive routing rules 3.4
Algorithm for Optimizing the List of Rules by Ranking Sorting An important parameter for the list of routing rules, in addition to the breadth of resource coverage and the size of the list, is also the arrangement of rules. To ensure the quality of service on the network, it is necessary to apply a sorting of the list of rules in order to place the least loaded routes in the top of the list. To solve this problem, the following algorithm was developed. Let's describe it.
Input: - A set R of tra c headers, passing through a virtual network device; - Non-con icting list of rules R = fRig ; i = 1; n; 1 Step 1: To assign each rule and each element of the set R, the weight by formula wi = kki , where ki - the number of tra c passes by the rule i; k total number of packets passing through the network. 2 Step 2: 3 if wn > w then go to Step 3; 4 else go to Step 5; 5 Step 3: Create a rule Rn 1, according to the current ow. 6 Step 4: Sort the set R by its value, place the most popular rule on the last place, assigning the corresponding metric go to Step 2. 7 Step 5: End.
Algorithm 2: Algorithm for optimizing the list of rules by ranking sorting
This algorithm does not reduce the size of the routing list, but allows you to make a more optimal list of rules.
Based on the proposed solution, the module implemented adaptive routing for a software-con gurable network. The software is implemented as a virtual network function based on the Open Platform for NFV (OPNFV), assembled as a Docker container.
Comparative analysis was carried out by comparing the results of traditional routing methods for a software-de ned network and a developed software module. Before the experiment, the developed software module was trained, and a comparable set of rules for traditional routing methods was developed, which allows for a correct comparison of the compared means. Within the framework of the study, work was evaluated under di erent loads for two key indicators: response time in the network; load the central processor on the NFV Router.
For carrying out of the load, test of scenarios. It includes 4 OpenFlow switches (2 HP 3500yl, 2 Netgear GSM7200), 8 computing nodes (32Gb RAM, 4 cores), 1 server (32Gb RAM, 8 cores) with OpenFlow controller and 1 server (32Gb RAM, 4 cores) for monitoring function. As a selected fat tree topology with three levels. Routers are connected with the speed of 1000 Mbit / s, and the computers are connected to the third level. In this infrastructure was prepared 100 virtual machines. Also we include one node that controls the data ows. We select random ve virtual machines (service host). Load is created using a specially designed tool - hping3. It allows you to generate packets of a certain type in certain directions, which makes possible on di erent values of the number of packets, and as a consequence. The experiment consists in the sequential measurement of the indicators with a consecutive increase in the intensity of tra c ows in single-load frames in the range from 20 to 350 Mbit / s.
Response time is one of the key parameters when analyzing network performance and network resources. It shows how much time passes from the moment the request is sent by the user until the service responds to the request. In accordance with the experiment rules, the response time from the network was removed.
The list of routing rules formed by the built module under load, including avalanche, gives a response time di erence of up to 20%, which indicates the e ectiveness of the approach developed in the work.
The load on the CPU of the virtual network device is an important parameter in ensuring the quality of service in the network supporting large data processing. This is an indication of whether the device is loaded, or ine ciently used. Based on the obtained results, we have that an optimized set of rules built using the approach considered in the study, has a reduced load on the rewall processor. 5
The study of the problems of route optimization and maintenance of quality of service in data transmission networks involved many scientists from di erent countries.
For example, authors of the publication [
In work [
To de ne rules for routing network tra c and ensuring quality of service,
we need to determine what types of tra c are present in a particular network.
To this goal, a number of researchers proposed approaches based on the
classication of tra c [
In a joint study of scientists Park J., Tyan H.-R., and Kuo C.-C., a new
scheme for classifying network tra c is proposed. The proposed approach does
not take into account the features of asymmetric routing and errors in modern
measuring instruments. To address these issues, the authors propose a two-stage,
scalable classi cation of tra c. At the rst stage, the functions for classi cation
are selected. At the second stage, the authors propose to use the method of
reducing the dimension of the classi ed sample. This will reduce resource intensity
and increase productivity [
Using multipart modi cation of the TCP (multipath TCP, MPTCP)
protocol in the data centers provides an increase in application performance. The
publication of Turkish scientists is devoted to the description of the architecture
using MPTCP based on PCS with the aim of guaranteeing customer service in
the issues of streaming video [
Amin Vahdat and its colleagues present design and experience with Google
Cloud Platform's network virtualization stack. Authors solutions take into
account a variety of factors including performance isolation among customer virtual
networks, scalability, rapid provisioning of large numbers of virtual hosts,
bandwidth, and latency. But they do not take into account the loud on the network
devices and QoS requirements in the network [
The researchers Lemeshko et al. in its works prepare the tensor model of
QoS-routing, where an accounting of the network state. They reduce the
QoSrouting problem to the solution of the optimization problem associated with
di erentiated maximization the probability of timely delivery of packet ows in
the network having a di erent service priority [
In the other work by Shuangyin Ren et al. prepared solution for
End-toEnd QoS guarantee algorithm for tra c routing based on deterministic network
calculus. They realized algorithm in Mininet with Open vSwitch as forwarding
switch and Ryu as the remote controller. The main idea of the solution is to run
Hierarchical Token Bucket queuing discipline to manage bandwidth resource.
The main problem with this solution is the lack of consideration of the load on
the communication channels at various times [
The analysis of scienti c sources on the topic of the study has shown that: a) so far, there are no e ective algorithmic solutions for planning tra c routing, application-oriented accounting topology of the computer system, and QoS requirements in communication tasks schemes;
b) the existing methods of data ow routing can be enhanced by taking into account the QoS requirements, but the need for fast solutions to analyze the current state of links in the network.
This demonstrates the novelty of the solutions o ered by the project. Thus, the development of new methods and algorithms to improve the e ciency of adaptive tra c routing in the cloud system infrastructure for Big data analysis is a crucial task. 6
The survey reviewed existing solutions in the eld of network routing. The architecture of the neural network was selected to achieve this goal, namely, a hybrid arti cial neural network consisting of two networks, namely a classi er based on a multilayer perceptron and a clustered based on the Kohonen network, the existing approaches, and algorithms for solving the problem of providing adaptive routing in the virtual network. In the framework of the study, a model for describing the rules for adaptive routing was developed, and on the basis of these model algorithms for automatically constructing and optimizing the list of routing rules were developed. In the projected architecture, an algorithm of con ict-free optimization was implemented, which produces the nal optimization by ranking and deducing the most frequently used routes, which allows improving the quality of service in the network. The algorithms and methods considered were implemented as a virtual network function for a software-de ned network.
During the tests, it was revealed that the approach proposed by this study gives an increase in two key parameters, namely, response time, on average by 20%, with an increase in load and load of the central process of the shielding device, by an average of 4.5% load growth. Thus, the developed approach is e ective for solving practical problems.