Figure 23 – uploaded by Lav Gupta

See full PDF downloadDownload figure

1etworks right to every host. The CloudLab set up created for this study is shown in Figure 5.9. international infrastructure from Internet2, so it has been possible to extend, software-definec — Figure 5 1etworks right to every host. The CloudLab set up created for this study is shown in Figure 5.9. international infrastructure from Internet2, so it has been possible to extend, software-definec

Related Figures (121)

Figure 4.1 OpenADN Multi-Cloud Management System

It chooses an initial set of clouds and acquires a pre-configured minimum number of virtual

Figure 4.3 Functions of OpenADN relevant to the bootstrap process backoff and repeats request checks for resources and launches workflow instances.

guest OS is Ubuntu 12.04 LT on all the VMs.

Figure 4.6 The Layered Profiling Model measure the absolute timing of events for carrying out the experimental study.

Table 4.1 Platform level execution data intensive.

Table 4.2 Time used for user and system activities on physical machines This simple profiling indicates the possibility of higher load on the CPU, because of potentially (unless otherwise indicated, all times are in seconds).

Figure 4.7 User, System and overall CPU time for OpenADN

Legend: nealls: the total number of calls, tottime: total time spent in the given function (excluding sub functions) seconds, percall: tottime divided by ncalls, cumtime: total time in this and all sub-functions seconds, Percall: cumtime divided by primitive calls, filename: data for each function between modules on different virtual hosts.

Legend: Hits: Number of times that line was executed, Time: Total execution time Per Hit: Average amount of execution time, % Time: Percentage of time spent on that line relative to the total amount of recorded time spent in the function, Line Contents: Actual source code. Figure 4.9 Extract of Statement Level Profile showing large sleep time 2000 users

the multi-threaded platform, thread level profiling was carried out while the program was in execution. A sample of concurrency profile is given in Figure 4.12.

‘workloads’ and the ‘experimental’ errors. The second column of Table 4.4 shows the sum of 7.66% and 31.35% more.

important for optimization regardless of the workloads.

effects related to functions and workloads (Table 4.6). The ‘t’ value used for the calculation of Table 4.5 The estimated y and the residuals

Table 4.6 Calculation of CI of Effects It can be seen that Cls of the functions are all significant. Also the means of a do not lie in the CI

Table 4.7 Performance of client connection for The Port initialization delays: This activity is required for connecting any node or service with he Port initialization delays: This activity is required for connecting any node or service with affected i.e., connection of the client and the http service to OpenADN are not hindered.

Figure 4.16 Setting up the pPort and overall reduction in the initialization time by 77.1% (Lines 206 and 211).

Total time for initialization 0.001938 seconds Tee A EO) Ghattlae ena thie Total time for initialization 0.002394 seconds Figure 4.18 Setting up the cPor

4.20 and 4.21 line number 125 For sPort the reduction was about 87% with the overall reduction by 40.7% as seen from Figures

Table 4.8 gives a summary of the time gained by optimization of initialization of various ports Total time for initialization 0.002041 seconds Figure 4.21 Setting up the sPort (after optimization udp tunnel port init. The concerned process initialized a communication port with the GC and

VNFs and SFC through the EMSs. Operation Support System/Business Support System (OSS/BSS) of the carrier manages the

Figure 5.2. Mapping service function chain to the multi- reconfigured by scaling up functions or even moving the VNFs to different clouds.

5.5.1 Information Available from Carriers and CSPs

Table 5.1 Symbols used Some of the important constraints subject to which the cost optimization is carried out are:

subset of least-cost clouds meeting the latency constraint. The complete algorithm is given ir

State at the time of planning deployment clouds, which give cost and latency below the stated thresholds, can be determined. eing able to provide service. When the states of the target clouds are known, the set of least-cost

Figure 5.5. (a) Traffic variation on Chicago-Seattle link

Figure 5.5 (b) Traffic variation on Los Angeles-San Jose Link

Figure 5.6. Comparison of generalization error with an integrated model and FPTV model

short-term changes in traffic as already discussed above. not been separately elaborated as it is based on the SVR model(s) refined for concept drift anc Algorithm 5.3 gives the details of random selection. The procedure PRI ENCY has EDICT LAT] set of clouds. The total cost of the selected clouds is checked against the current minimum cost, gnore any of the clouds in the space N. These conditions are met in our implementation

relationship with the algorithms discussed. The placement strategy described above has been implemented in a placement framework called

Table 5.2 Categorization of server resources 5.10.2 Selection of Features for Training the Prediction Models Considering the importance of the selection of predictor variables, due attention was given to this storage. In our experiments, the categories defined are as shown in Table 5.2.

Table 5.3 Predictor variables and the output label 5.10.3 Obtaining Training Datasets We were cognizant of the fact that if a model has been trained with the adequate, realistic obtained good compensation of concept drift with four windows as indicated by the results.

The total time spent by voice and data packets in the network can follow any distribution. queueing delays on the link. Intercloud simulation was carried out covering all significant delays

Table 5.4 An extract of the integrated training dataset 5.10.6 Selection of the Machine Learning Model training sets is shown in Table 5.4. There are quite a few AI techniques, involving machine learning, that are potentially applicable

unlimited. The number of iterations or number of trees is set as 100. when new examples are introduced takes time. The maximum depth of each tree has been set at which is not linearly separable. Parameter values that obtained for good results are C=200,

A comparative plot of training and test error ratios (defined as prediction_error/acutal_latency) dataset. These can be overcome by curating the training dataset. is given in Figure 5.10. It can be seen that the model training errors are low and it generalizes be slightly higher than the training errors. A lower test error may indicate overfitting or biases in the

5.10.8 Refinement of Latency Prediction by Compensating Concept Drift The FTVP method for handling the concept drift in telecommunication traffic was presented in

Figure 5.11 Extract of FTVP windows divided into four windows. A sample from each of these is given in Figure 5.11.

Table 5.8 Probability distribution parameters in different windows SVR with separate window models gives much better predictions on new data-points falling in those

in Figure 5.12 (a) through (h)

errors show an improved model. Table 5.11 Performance after online updation

In another trial of 5000 experiments, 50 each with the number of clouds increasing from 10 to

Figure 5.16. Latencies achieved by randomized and M-Sequential in the number of iterations shown in Figure 5.15 Figure 5.15. Number of iterations required by randomized and M-Sequential to achieve latency below the threshold

other works we examined the work done in [148]. The two methods have been performed under

Figure 5.18 Placement time reported in [148] network functions per request or a total of 3000 functions) take about 1000 seconds.

Figure 5.19. Average placement time reported in [149] functions/instances JoraNFV will take 4285 seconds (Figure 5.19). while SA takes about 2000 sec. Even if we assume a linear increase in time taken, for 3000

Figure 5.21. Acceptance % reported in [153] end-to-end delay and resources allocation cost. A comparison of successful requests is given in

Figure 6.1 A carrier’s broadband service network

a) Virtual Network Service (VNS) functions in a top-down manner. From the illustration of a complex network service in Figure 6.1, we abstract a subset to

b) Service Function Chains (SFCs) A carrier may obtain resources from multiple cloud service providers for the reasons discussed

resources to form one or more SFCs [165] [166].

1) Virtual Machines (VMs) and Network Function Virtualization Infrastructure (NFVD sed to deploy VNFs. This is also referred to as a NFVI Point of Presence (NFVI-PoP). The

Table 6.1 Examples of commercially available VNFs shows a VNF with its |] Element Management System (1 EMS) and interfaces to the rest of the

Figure 6.6. Network function virtualization infrastructure e) Routing Protocols for SFC connectivity may be possible through the designated switch and routing devices [165].

6.4.1 Virtual Infrastructure Manager (VIM)

6.5 Comparison of Competitive Network Service Orchestration Offerings

information is not available to adjudge the product on the corresponding criterion.

6.6.3 The FP Problem Statement The FP problem of the carrier networks can be defined as follows:

Table 6.5 List of markers for different carrier services related to mobile, fixed and broadband networks are given in Table 6.5. able to extract relevant features automatically, without human intervention. Some of the markers

6.8.2 AI Techniques used in the Framework values (where applicable), from an actual network [174] are given in Table 6.6. A number of AI techniques, involving machine learning and deep learning, were tested and the

Table 6.7 summarizes the machine learning and deep learning techniques useful for NFV-Cloud

6.8.3 Design and Implementation of the HY PERVINES Framework

In the training mode, the data is curated and partitioned into training and test datasets. Shallow

responsibilities and ill-defined interfaces of the management platforms. the Detection and Localization functions with an optional inclusion of the data pre-processing TYPER-VINES obtains markers from multiple sources, it mitigates the problem of overlapping module. We discuss the important details of the framework below.

particularly important, as the presence of alarms does not always indicate a fault.

Table 6.8 Coarse (Layer!) and fine (Layer 2) categorization of faults the maximum margin can be classified with f(x) = arg max; fi(x), i.e., choose the class that classifies the example with

parameters. Let us discuss a little more about the manifested and impending faults. maintenance is relied on to catch issues early. In our framework, localization of impending faults procedure impending localization (line 17) calls the deep learning model with the required

stacking a number of autoencoders and carefully designing the hidden layers.

Figure 6.14 Stacked sparse autoencoders minimize the mean square error in the labeled test datasets [29]. tuned using back-propagation and simultaneous adjustment of weights of all the layers |

Table 6.9 (a) Training dataset The test dataset has “id” and the /ocation for which severity has to be predicted

Table 6.10 List of features from network fault dataset complete set of features from featurel to feature386) is shown for compactness. As part o indicating the intensity of the warning and fault_severity ranging from 0-2 as explained before.

this study, a correlation threshold of 23% was found to improve accuracy.

With this dataset, SVM, on the whole, performs better than ADT and RF giving > 95.49% comparatively better than RF. In each case, 10% cross-validation was used.

Table 6.14 Detection Level | performance comparative baseline and the accepted model (SVM in this case) results are shown in Table 6.14

For Level-2 detection, we have chosen One-R as the baseline algorithm. The accuracy of ow framework is 13.03% better for ‘impending’ faults and 5.97% better for ‘manifested’ faults, Kernel works well as can be seen from Table 6.15.

Figure 6.16. Detection Level 2 effectiveness compared to baseline

At Layer-1, we chose the sequential minimal optimization (SMO) multi-class support vector Table 6.16 Localization Performance the accuracy of Layer-1 localization is 97%, which is a substantial improvement over the baseline

Figure 6.18 Localization Layer 2 effectiveness compared to baseline 6.9.4 Localization of Impending FP Issues

sparsity proportion (SP). SP gives the proportion of training examples a neuron reacts to. A low

Figure 6.21 Localization of impending fault stacked autoencoder and SVM (baseline) severity levels of the impending faults

Figure 7.1 Overall layout of the loT-Cloud based healthcare the next generation healthcare service in Figure 7.1. 7.3.1 The Healthcare Network Domains

the network. A brief discussion on the important constituent of various zones is given below: gateway in the layout diagram of Figure 7.1. This zone is composed of devices like sensors,

Figure 7.3 The security architecture of the next generation healthcare Figure 7.3 in a simplified form.

Figure 7.4 Trust boundaries between zones Figure 7.4. 7.4.1 Attack Surfaces

Table 7.1 Suitability of deep learning for IoT healthcare networks as the data is passed through the layers of the deep structure of the networks, the shown in Table 7.1.

Figure 7.5 Hierarchical autoencoder based system The dataflow coming from the IoT domain is passed through the autoencoder of the edge clouds intrusion system is given in Figure 7.5.

the edge models are merged to create public cloud model. way the training of layers at the edge can be used in the public cloud. Figure 7.6 illustrates how

Table 7.2 The feature set of the BoT-IoT dataset

Figure 7.7 Healthcare testbed for dataset generation The IoT domain: Consists of an Arduino Mega based microcontroller and gateway for health IoT sensors. Because of the absence of any Wi-Fi or Ethernet post on this microcontroller, an

Figure 7.8 Extract of the Training Dataset The normal and attack datasets have the features given in Table 7.3. An extract of the

autoencoder layers. The other edge cloud models are created similarly.

Train-loss = 4.697, Test-loss = 17.595, Train-accuracy = 0.993, Test-accuracy = 0.995

Train-loss = 8.526, Test-loss = 8.661, Train-accuracy = 0.986, Test-accuracy = 0.985

cloud neural network models created ab-initio with the models created by reusing trained layers from edge cloud neural networks.

Train-loss = 4.517, Test-loss = 6.151 Train-accuracy = 0.989, Test-accuracy = 0.990 a) Public cloud merged layers cloud while Figure 11 b) shows results for the merged model.

Train-loss = 2.1575, Test-loss = 4.1718, Train-accuracy = 0.992, Test-accuracy = 0.993 b) Public cloud merged layers (cross-trained) Figure 7.11 Public cloud merged model training and testing

trained in the edge clouds.

To see the performance of the model in filtering out the attack cases, we used the attack data

The random runs generally give low false positives and the accuracy ranges from about 95% to for several runs is given in Table 7.4. value that decides whether the data has been affected by malicious activity. The confusion matrix Summary and Future Directions

Figure 7.14 Learning Performance and Explainability of current and future systems [DARPA]