Virtual Network Function Scaling

Performance Evaluation 

The performance of our proposed ILP based is assessed through extensive simulations. Simulator settings and performance evaluation metrics used for evaluation and comparison purposes are presented. The ILP is compared to the Greedy heuristic. The evaluation focuses on the gains reconfiguration (just to tidy up the network) and scaling (during rising demands) provide to the stakeholders by analyzing the proportion of rejected (failed) adaptation attempts.

Simulation Environment 

Our simulations are based on a realistic topology as well as extensive simulations for which it is possible to evaluate in depth the scalability of the algorithms using larger infrastructures and request sizes. A 2.50 GHz Quad Core server with 6 GBytes of available RAM is used for the performance evaluation and comparison of the proposed algorithms. For the realistic topology and the extensive simulations, the VNF-FG requests are generated using a Poisson process with an average arrival rate of 5 requests per 100 time units. The lifetime of each request follows an exponential distribution with a mean of 1000 time units. The Germany50 network topology [85] is used for the first assessment (with 50 nodes). This topology is defined by the German National Research and Education Network (DFN). The capacity of physical nodes and links are generated randomly in the [50, 100] interval. The size of the VNF-FG requests is set to 5 nodes. The GT-ITM [86] tool is used to generate the requested VNF-FG topologies. The initial VNF-FG computing and bandwidth requirements are set to 10. Scaling in CPU is fixed to 20 with one scaling request at a time per VNF-FG. To assess the scalability of our proposed algorithms, we generate using the GT-ITM tool a network topology with 100 nodes and a connectivity of 0.2 (or 20%). The physical resource capacities (i.e., CPU and bandwidth) are also drawn randomly in the [50, 100] interval. The VNF-FG request sizes vary between 3 and 15 nodes. The required CPU for each VNF in the VNF-FG is set to 10 units. The required bandwidth between two VNFs, to ensure communication between them, is set also to 10 units. The connectivity between nodes in the VNF-FG is set to 0.3. The ILP algorithm is evaluated for three values of the migration penalty P enu from 10 to 1000 units in order to tune the penalty according to the needs and provider priority as well as to measure its effect on performance. The migration penalty (P enu in the second term of the objective function in Equation 3.3.8) is used to control the re-mapping process and to especially trade-off “in node scaling” with “migration to other hosts”. The migration penalties in the performance evaluation correspond to the reported ILP10, ILP100 and ILP1000 results. The performance of the ILP for these penalty values (10, 100 and 1000) are also compared to the Greedy algorithm, to assess the effectiveness and usability of this migration. For the realistic topology and the simulation, 1000 VNF-FG requests are generated and a single VNF scaling request is triggered for each and every generated request. Since, we are focussing on scaling and adaptation of already embedded VNF-FGs, we used the heuristic approach of [58] to generate the initial VNF-FG mapping and initialize the assessment runs.

Performance Metrics 

The metrics used for the performance evaluation are described in this section that reports also the results obtained using both the realistic topology and the extensive simulations for the extended performance assessment.

• Successful scalings represents the number of VNF scaling requests that have been accomplished. Indeed, due to physical resources limitation the algorithm may reject some scaling requests. From the provider point of view, this number should be maximized in order to improve the global revenue.
• Scaling ratio is the ratio of successful scalings defined by the number of successful scalings to the accepted VNF-FG requests. The number of accepted VNF-FG requests depends on the initial provisioning and on how the scaling algorithm deals with scaling requests. Since we use as initial mapping the same basis for all algorithms, the scaling ratio reflects their relative efficiency and enables their comparison. Virtual Network Function Scaling 
• Number of migrated VNFs is the number of the scaled VNFs through migration. Migration involves a temporary service interruption until the concerned VNF is activated in the new physical host.
• The ratio of migrated VNFs is the ratio of the number of migrated VNFs to the number of successful scalings. This metric measures the proportion of adaptations accomplished through migration that providers prefer to minimize to avoid or limit the service interruptions due to migrations. Minimizing this measure reduces disruptions to applications.
• Execution time is a decisive measure in order to assess the scalability of the algorithms. Service providers prefer efficient and rapid algorithms in order to quickly serve clients.