IEEE ICC 2012 - Dependability Assessment of Virtualized Networks

Stenio Fernandes, Eduardo Tavares, Marcelo Santos, Victor Lira, Paulo Maciel

Federal University of Pernambuco (UFPE)Center for Informatics

Recife, Brazil

Dependability Assessment of Virtualized Networks

Outline

Motivation, Problem Statement, and Proposal Related Work Technical Background Hierarchical Dependability Modeling and

Evaluation Dependability Assessment of VNs Contributions and Future Work

MOTIVATION, PROBLEM STATEMENT, AND PROPOSAL

Motivation (1/3)

Network Virtualization is a paradigm shift to allow highly flexible networks deployment

Virtual Networks (VN) – have intrinsic dynamic aspects

It allows operators to have on-demand negotiation of a variety of services

Important properties: concurrent use of the underlying resources, along with router, host, and link isolation and abstraction

– resources reuse is performed through appropriate resource allocation and partitioning techniques

Motivation (2/3)

Network Virtualization management strategies – rely on dynamic resource allocation mechanisms for

deploying efficient high-performance VNs Goal: achieve efficient resource allocation of the physical

network infrastructure– heuristic approaches due to its NP-hardness nature

– Efficient partitioning and allocation of network resources is the fundamental issue to be tackled

Physical Networks

Composed Network - Virtual

Motivation (3/3)

However, from the point of view of the end-user – a Service Provider or any entity that wants to build VN

to offer services there is still a missing point:

– What are the risks associated to a certain VN?

Problem statement

Argument & hypotheses: – risks are inherent to virtualized infrastructures since

the underlying physical network components are failure-prone E.g., subject to hardware and software components

failures– Understanding Network Failures in Data Centers:

Measurement, Analysis, and Implications, SIGCOMM 2011 – A first look at problems in the cloud. USENIX HotCloud 2010

– Risk is a crucial factor to the establishment of Service Level Agreements (SLA) between NV engineering and business players

Problem statement

Risk evaluation and analysis, from assessment of dependability attributes, can quantify and give concrete measures to be used for network management and control tasks

Risk evaluation must be taken into account when formulating an optimization problem for resource allocation and provisioning of components at the physical network

Proposal

This paper proposes and evaluates a method to estimate dependability attributes (risks) in virtual network environments, – It adopts an hierarchical methodology to mitigate the

complexity of representing large VNs Reliability Block Diagram (RBD) Stochastic Petri Nets (SPN)

Assessment of dependability attributes could be adopted as a critical factor for accurate SLA contracts

RELATED WORK

Related Work

Xia et al. tackle the problem of resource provisioning in the context of routing in optical Wavelength-Division Multiplexing (WDM) mesh networks– Risk-Aware Provisioning scheme that elegantly minimizes the

probability of SLA violation "Risk-Aware Provisioning for Optical WDM Mesh Networks," Networking,

IEEE/ACM Transactions on, June 2011 Sun et al. proposes a cloud dependability model using

System-level Virtualization (CDSV), which adopts quantitative metrics to evaluate the dependability – They focus on cloud security and evaluate the impact of dependability

properties of the virtualized components at system-level "A Dependability Model to Enhance Security of Cloud Environment

Using System-Level Virtualization Techniques," 1st Conference on Pervasive on Computing Signal Processing and Applications (PCSPA), 2010

Related Work

Techniques for assessing dependability attributes have been evaluated in virtual computing systems. – SPN and Markov models have been adopted to assess

them in VMs and Oses. Koslovski et al. takes into account reliability only

support in virtual networks– it has a general view on nodes and links at the physical

infrastructure– it does not take into account the hierarchical nature of real

systems, Composed of virtual machines, disks, operating systems, etc.

– "Reliability Support in Virtual Infrastructures”, IEEE CloudCom 2010

Related Work

In general– Simplified views

Specific to components, sub-systems, etc OR Consider only a direct mapping between the

physical infrastructure and a given VN– little effort on research studies that provide

dependability measures for risk assessment They could be adopted as input for resource

allocation algorithms and provisioning techniques

TECHNICAL BACKGROUND

Technical Background

Dependability of a system can be understood as the ability to deliver a set of services that can be justifiably trusted– It is also related to fault tolerance, availability, and

reliability disciplines Dependability metrics can be calculated by

– Combinatorial Models Reliability Block Diagrams (RBD) and Fault Trees

– State-based stochastic models Markov chains and Stochastic Petri Nets (SPN)


Some dependability metrics – Availability (A) of a given device, component, or system

it is related to its uptime and downtime Time to Failure (TTF) or Time to Repair (TTR) Mean Time to Failure (MTTF) and Mean Time To Repair

(MTTR) – Steady-state availability (A) may be represented by the

MTTF and MTTR, as:


MTTF can be computed considering the system reliability (R) as

Exponential, Erlang, and Hyperexponential distributions are commonly adopted for representing TTFs and TTR– i.e., adoption of semi-markovian solution methods

HIERARCHICAL DEPENDABILITY MODELLING AND EVALUATION

Hierarchical Dependability modelling and evaluation

Proposed methodology for dependability evaluation of virtualized networks

Three steps

System specification

Subsystem model

generation

System model

construction


• information concerning the dependences of VNs and possible mutual impacts, such as Common Mode Failure (CMF)

• information related to the TTF of each component or sub-components and the respective TTR

System specification


• the system may be represented either by one model or split into smaller models that comprise system parts (i.e., subsystems).

• Such an approach mitigates possible state space size explosion for large and detailed models

Subsystem model generation


• intermediate results are combined into a higher level model using the most suitable representation• For instance, physical nodes are

initially represented by a RBD model (using series composition) and the obtained results are adopted into a SPN model.

• Final model is then constructed by using the metrics obtained in previous activity and, lastly, such a model is evaluated.

System model construction


Proposed method provides the basis for obtaining the dependability metrics and for evaluating quantitative properties

It utilizes Mercury/ASTRO environment for modeling and evaluating dependability models

– Tools available to academics (under request)

DEPENDABILITY ASSESSMENT OF VIRTUAL NETWORKS

Dependability Assessment of VNs

Evaluation Methodology

• Generation of several VNs requests that must be allocated on the top of a common physical network

• For each new allocated VN, we assess dependability metrics for each system and subsystem in the physical and virtual network

• We assume that dependability metrics are known for each component of the network, including their subsystems. • Information from real measurements and data are

available in the literature• Depending on the chosen model, dependability metrics

may change for each new VN allocation


Virtual Network Topology Generation (R-ViNE)– the substrate network topologies are randomly generated

using the GT-ITM tool; – Pairs of nodes are randomly attached with probability 0.5;

500 VN requests during the simulation time (50,000 time units) in a network substrate with 50 nodes. – VN requests follow a Poisson process with mean λ = 4

(average of 4 VNs per 100 time units);– Each VN follows an exponential distribution for its lifetime

with λ = 1000 (i.e., an average of 1000 time units); – For each request, the number of virtual nodes per VN

follows a uniform distribution in the interval [2, 10].


Case Study mapping algorithm proposed in [3]

– "Virtual Network Embedding with Coordinated Node and Link Mapping”, IEEE INFOCOM 2009

– The algorithm provides VN allocations in an infrastructure provider satisfying CPU, link, and other constraints.

– It does not assume dependability issues, which may impact the feasibility of a given allocated VN

We applied the resource allocation algorithm to evaluate the dependability features for each allocated VN


Case study (cont.)– demonstrate the estimation of point availability (i.e.,

availability at a time t) and reliability– assuming independent allocations and common mode

failure (CMF) we assume that the components are connected via series

composition– if a component fails, the virtualized network fails


Typical MTTFs and MTTRs

Node MTTF (h) MTTR (h)

CPU 2500000 1

Hard Disk 200000 1

Memory 480000 1

Network Interface Card 6200000 1

Operating Systems 1440 2

Virtual Machines (VM) 2880 2

VM Monitor 2880 2

Switch/Router 320000 1

Optical Link 19996 12


VN net0 has a lower availability level, when CMF is assumed

the algorithm could avoid overload in some links and nodes with smaller MTTFs


Availability measures for the sampled VNs are very similar– In more complex environments, dispersion metrics can

vary significantly

Extensions to the resource allocation algorithm

Mapping algorithm might have to take into account one or more dependability measures– To meet strict requirements

For instance, a Service Provider can require an availability of 0.95 and minimum reliability of 0.99 during the lifetime of a certain VN.

Allocation alternatives– to minimize the impact on availability and reliability of

previously defined VNs– to improve the dependability measures of a new VN

allocation

CONTRIBUTIONS AND FUTURE WORK

Contributions and Future Work

Contributions– an approach for dependability modeling and evaluation

of virtual networks using a hybrid modeling technique that considers representative combinatorial and state-based models.

– The proposed approach provides a basis for estimating dependability metrics, such as reliability and availability, which we consider important for heuristics dealing with resource allocation in VNs

Contributions and Future Work

Future Work– analysis of fault-tolerant techniques to improve

dependability levels when the ordinary components are not able to achieve the

required service level– formulate an efficient optimization model in the way that

dependability metrics can be handled as range of values

IEEE ICC 2012 - Dependability Assessment of Virtualized Networks

Technology

dependability measures

network virtual

cloud dependability

network failures

network management

virtual network environments

related work techniques

risk assessment