Constrainedbased QoSAware Service Adaptation

1
Constrained-based QoS-Aware Service Adaptation
The Service and Software Architectures,
Infrastructures and Engineering (SSAIE) Summer
School

• Kyriakos Kritikos
• Post-doc Researcher
• Politecnico di Milano

2
Outline

• Polimis IS group research
• Introduction
• Background
• Constrained-based QoS-Aware Service Adaptation
  (CQSA)

3
IS Group Research (I)

• Main research on services
• Mapping KPIs on QoS and their assessment in
  service-based processes
• Expressing KPIs in terms of QoS
• Measuring and assessment of KPIs
• Automated service negotiation
• Expressing negotiation concepts (e.g.
  capabilities, cost models, strategies) through
  ontologies
• Ontologies rules -gt reasoning
• Protocol discovery delegation 1

4
IS Group Research (II)

• Main research on services
• Energy-aware design of service-based processes
• GPIs like energy efficiency and consumption
• Constrained-based Service Concretization
• First two research topics performed within S-Cube
• More information at
• http//home.dei.polimi.it/pernici/ws-research.html

5
Introduction (I)

• Facts
• Services operate in highly dynamic environments
• Variable execution context
• Service should change its behavior
• Runtime conditions differ from presupposed design
  time ones
• Service should be able to execute

6
Introduction (II)

• Requirement Self-Adaptation
• Services have to be
• Flexible timely-respond to changes in their
  context or user preferences and context (design)
• Adaptable detect or even predict system
  failures, react on them and compensate for
  deviations in functionality or quality
  (self-healing mechanisms)
• Our work focuses on quality aspect of service
  adaptation

7
Introduction (III)

• Quality-aware Service Adaptation
• Agreed Service Level (SL) in Service Level
  Agreement (SLA) must be met
• SL set of quality guarantees (QG) on quality
  metrics
• QG is monitored evaluated
• If QG is violated, then service must be adapted
  by executing some actions
• E.g. service renegotiation, SLA termination,
  adding more resources to service, nothing
  (penalty will be paid), reselection for composite
  service

8
Introduction (IV)

• Research has been focused on re-active service
  adaptation
• Faulty services are executed resulting in loss of
  money and executing additional activities
• Adaptation activities take time and reduce system
  performance
• Monitoring events may arrive too late

9
Introduction (V)

• Solution -gt Pro-active service adaptation
• Predict service quality
• Pro-actively respond to predicted deviations
• Main advantage
• Predict deviation and react before it occurs
• Main disadvantage
• Prediction is not always accurate
• False positives
• Usually reduces system performance

10
Introduction (VI)

• End-to-End Quality Provision and SLA Conformance
  (WP-JRA-1.3 S-CUBE)

11
Background

• Facts
• SOA is based on the SOA triangle
• Service registry implementations have failed
• Service Discovery Problems
• Syntactic approaches in WS registries (WSDL)
• No means to select among functionally equivalent
  WSs

12
Background 1st Problem (I)

• 1st Problem
• Textual or XML-based descriptions lead to
  accuracy problems
• E.g. User requests for a function add(X,Y)
  adding two reals, while service provides the same
  function that adds two integers
• Registry will return this service description to
  the user

13
Introduction 1st Problem (II)

• Solution
• Use of ontologies that provide a
  machine-understandable and unambiguous vocabulary
  of terms
• Conceptualize a specific application domain
• Map each service object (i.e. WSDL element) to an
  ontology concept
• Use Semantic Web techniques in registry to infer
  the similarity or equivalence of two terms (user
  and service WSDL files)
• Increased accuracy in service discovery

14
Introduction 2nd Problem (I)

• Solution to second problem
• QoS set of non-functional properties
• Each property either measured by QoS metrics or
  is unmeasurable
• QoS spec is a combination of constraints on these
  QoS metrics and unmeasurable properties

15
Introduction 2nd Problem (II)

• Solution to second problem
• QoS in WS description and discovery
• Apart from the functional, the QoS aspect of
  services is described
• QoS specifications (offers and demands) are
  matched based on a matchmaking metric
• The best service is selected based on user
  weights on QoS metrics and unmeasurable properties

16
Introduction 2nd Problem (III)

• Domain-independent QoS attributes
• Response time (M), execution time (M),
  availability (M), reliability (M), data
  encryption (U), reputation (M), price (U)
• Domain-dependent QoS attributes
• Traffic Monitoring application domain Refresh
  time (M), covered area (U), routes set (U),
  detail level (U), accuracy (M), completeness (M),
  validity (M), timeliness (M), coverage (M)
• M-gtMeasurable, U-gtUnmeasurable

17
Introduction 2nd Problem (IV)

• QoS spec example

18
Introduction 2nd Problem (V)

• Matchmaking Example
• Offer 1
• Offer 2
• Demand
• Both offers respect the constraints of the demand
• Selection example
• User Weights
• Offer 1
• Offer 2
• Offer 2 is selected for the user

19
Introduction CP (I)

• Constraint Programming (CP)
• Solves Constraint Satisfaction Problems (CSPs)
  V,D,C, Vset of variables, Dset of
  domains, Cset of constraints
• Solution to CSP is an assignment of every v in V
  from the domain of v that does not violate any of
  the constraints of C
• Solution space of CSP is the set of all possible
  solutions.

20
Introduction CP (II)

• A CSP is satisfiable if its solution space is not
  empty
• E.g. CSP (x,y,0..2,0..2,xlty,xgt0) is
  satisfiable as it contains the sole solution
  x-gt1,y-gt2.
• When an objective function over V is used to find
  the solution that minimizes its expression, then
  the CSP is actually a Constraint Satisfaction
  Optimization Problem (CSOP)

21
Introduction CP (III)

• Advantages
• non-linear constraints
• logical combination of constraints
• small domain integer variables
• close representation to physical problem
• Disadvantages
• use of real variables under research
• solving time increases with increase of domain of
  variables
• CP solvers do no deal well with optimization
  problems and do not propagate well with
  disjunctive constraints

22
Constrained-Based QoS-aware Service Discovery
(CQSD)

• Main Approach followed
• Service providers and requesters express their
  QoS spec in a common language
• These QoS specifications are transformed to
  different constraint descriptions
• These constraint descriptions are matched based
  on a specific matchmaking metric
• QoS offers are categorized into two categories
  fail and match
• QoS offers are ranked based on selection algorithm

23
Constrained-based QoS-aware Service Matchmaking
(CQSM)

• Matchmaking efforts fail
• Face accuracy problems
• Do not use correct matchmaking metrics
• Do not offer advanced categorization of results
• Offer valueless results for over-constrained
  demands

24
Incorrect Matchmaking Metrics (I)

• Matching (Degwekar et. al. 2004)
• Conformance (Cortes et. al. 2005)

O1
Better solutions
D
O2
O3
O4
25
Incorrect Matchmaking Metrics (II)

• Example
• Two offers O1 and O2 and one demand D
• X measures Availability (positively monotonic)
  and has value type (0.0,1.0)
• Offer O1 has 0.9ltXlt0.96, Offer O2 has
  0.96ltXlt0.9999 and demand D has 0.95ltXlt0.999.
  
• Matching metric Both offers match the demand
• Conformance metric None of the offers match the
  demand
• Reality Offer O2 should match with the demand D

26
Advanced Categorization of Results (I)

• Matchmaking algorithms return only two types of
  results
• What happens when only fail results exist?
• Ideal there must be 4 cases
• Case 1 Super demands space is subspace or
  lower than offers space
• Case 2 Match offers space is subspace of
  demands space
• Case 3 Partial demands and offers spaces
  intersect
• Case 4 Fail demands and offers spaces do not
  intersect and demand space is higher

27
Advanced Categorization of Results (II)

• super, match, and partial results are useful for
  service discovery and negotiation
• partial results are also useful for optimization
  i.e relaxing user constraints when super and
  match results are not present
• fail results are not useful at all!

28
Advanced Categorization of Results (III)
Better solutions
Super O1
D
Exact O2
Partial O3
Fail O4
29
Solution Matchmaking Metric (I)

• Conditional conformance an offer O matches a
  demand D when its CSP PO has solutions that are
  either contained in the solution space of the CSP
  PD of the demand or are better than the demands
  solutions.
• Prefer better solutions only for insensitive
  metrics

30
Solution Matchmaking Metric (II)

• By analyzing conformance, we have
• sat(CSPO ? ?CSPD) false?
• sat(CSPO ? ?(con1D ? con2D ? ? conmD) false?
• sat(CSPO ? (?con1D ? ?con2D ? ? ?conmD)
  false?
• sat((CSPO ? ?con1D) ? (CSPO ? ?con2D) ? ? (CSPO
  ? ?conmD)) false?
• sat(CSPO ? ?con1D) sat(CSPO ? ?con2D)
  sat(CSPO ? ?conmD) false
• So if demand D has m constraints, then we have to
  solve m CSPs
• If all CSPs are unsatisfiable, then offer O is
  conformant to demand D
• Trick take only into account the interesting
  constraints of insensitive metrics only

31
Solution Matchmaking Metric (III)

• Sensitive metric
• both low and upper bounds are interesting, they
  matter the user
• E.g. response time
• Insensitive metric
• Only low quality values matter the user
• Only low or upper bounds are interesting, for
  posit. mon. or neg. mon. metrics, respectively
• E.g. for availability only the low bound is
  interesting

32
Solution Matchmaking Metric (IV)

• Example
• Offer 0.96ltAvaillt0.9999
• Demand 0.95ltAvaillt0.999
• There is a match because we do not take into
  account the upper bound for availability

33
Solution CQSM

• Developed CP and LP algorithms of conditional
  conformance metric
• Still no advanced categorization of results,
  optimization
• Proposed two new algorithms unary and
  optimization one
• Unary 2 (wrt. the constraints of the demand) is
  like exhaustive algorithms
• Optimization 2 is quicker but has precision
  problems

34
Unary Algorithm (I)

• Composed of 2 steps
• matchmaking
• cons. relaxation
• Metrics are associated with weights
• Matchmaking step
• Solve all possible CSPs of an offer and demand
• If all interesting CSPs are inconsistent then
• If a non-interesting CSP is inconsistent, the
  offer is super match
• Otherwise, it is an exact match

35
Unary algorithm (II)

• Matchmaking step (cont.)
• If some interesting CSPs are consistent then
  offer is a partial match, weight counter updated
• If all are consistent, then offer is a fail match
• Constraint relaxation step
• it is executed, if no super or exact matches
  exist
• All results are sorted (decreasing) according to
  weight counter and the best ones promoted

36
Unary Algorithm (III)

• Example
• Assume alignment process produces four CSPs
  having the following definitions
• X1 (0.0,86400.0, X2 (0,100000, X3
  (0.0,1.0)
• CSP1 X1 ? 10.0, X2 ? 100, X2 ? 50, X3 ? 0.9
• CSP2 X1/60 ? 0.08, X2 ? 50, X2 ? 40, X3 ? 0.95
• CSP3 X1/60 ? 0.06, X2 ? 40, X2 ? 30, 100?X3 ? 98
• CSPD X1 ? 15.0, X2 ? 40, 100?X3 ? 99
• X1 measures response time
• X2 measures throughput
• X3 measures availability

37
Unary Algorithm (IV)

• Example (cont.)
• Matchmaking step checks satisfiability of 9 CSPs
  and produces 4 lists SuperExactFailed,
  Partial (O1,CSP1,1,10.0,100?X3 ? 99),
  (O2,CSP2,1,10.0,100?X3 ? 99), (O3,CSP3,2,20.0,X2
  ? 40,100?X3 ? 99)
• Constraint relaxation process sorts the partial
  list and produces 2 lists Exact(O1,CSP1),
  (O2,CSP2), X3, Partial (O3,CSP3, X2,,X3)

38
CP-based Matchmaking (I)

• It is better than LP when
• There are non-linear expressions in constraints
• Non-linearity usefulness
• Express dependencies between quality metrics
• E.g. availability is inverse proportional to
  response time
• Express cost functions
• E.g.

39
CP-based matchmaking (II)

• Examined Dynamic CP
• Chose re-active and more specifically reasoning
  reuse approaches
• Concluded explanation-based CP (eCP) is the best
• Implemented eCP versions 3 of conditional
  conformance and unary algorithm with Choco

40
(No Transcript)
41
CQSA Background (I)

• In literature, service quality is established in
  SLA
• SLA-gtoutcome of service negotiation executed
  after service discovery
• SLA comprises of constraints on metrics measuring
  quality attributes (i.e SLOs)

42
CQSA Background (II)

• In 4, eCP is used to find inconsistencies in
  SLA definition and the set of conflicting
  constraints
• By adopting the constraint approach, SLA
  violation detection is easy
• If coupled with monitoring
• Same goes for prediction of SLA violations

43
CQSA Background (III)

• However, in literature there are two main
  problems regarding service monitoring and quality
  prediction
• Single values are monitored predicted
• Confidence in prediction is low
• Monitored values are not always accurate
• Quality attributes are supposed to be independent
  of each other
• Wrong! Values on some attributes depend on each
  other (e.g. response time availability)

44
CQSA Proposed Approach (I)

• Use range constraints for predictions of quality
  metrics
• Confidence level will be higher than equality
  constraints (weight)
• Express dependencies between quality attributes
  with constraints
• Both quantitative and qualitative dependencies
  can be modeled in this way
• Qualitative can be characterized by a stochastic
  quantitative model (e.g. function derived from
  empirical data using regression)

45
CQSA Proposed Approach (II)

• The set of dependency constraints D can be used
  for validating correctness of monitoring or
  prediction
• By solving a CSP constructed from D
  predicted/monitored range of values
• The previous set plus another set of (probably
  unary) constraints U constitute the service SLA
  (UD)

46
CQSA Proposed Approach (III)

• The whole set of constraints (UD) composes the
  initial state of the service
• Each prediction (executed in specific time
  frequency) is a possible state of the system
• Each predicted state is compared with the initial
  one
• So quality-based service matchmaking technique is
  reused here
• Initial state -gt Demand CSP
• Predicted states -gt Offer CSPs

47
CQSA Proposed Approach (IV)
Initial State
Predicted State 1
Predicted State N
RTlt5 AVgt0.99 RTf(AV)
RTlt6 AVgt0.98
RTlt4 AVgt0.99
........
Dep. State
Mon. State N
Mon. State 1
THf(RT) RTf(AV)
5ltRTclt6
6ltRTlt7
48
CQSA Proposed Approach (V)

• How do we detect a violation?
• For each quality metric
• If both its range constraints violate initial
  service state
• If confidence level gt threshold then sure
  violation case
• Else possible violation case
• If one range constraint violates initial service
  state
• If confidence level gt threshold then maybe
  violation case
• Else unlikely violation case

49
CQSA Proposed Approach (VI)

• Adaptation strategy depends on number of
  different types of cases (and their combination)
• E.g. If one ore more sure violation cases
  detected, then either the service gets more
  system resources or service provider renegotiates
  with requester to relax appropriate SLA
  constraints
• E.g. If only unlikely violation cases detected,
  then time frequency of prediction increases and
  the services system is warned

50
CQSA - Architecture
Prediction Engine
Adaptation Engine
Execution Log
Assessment
Monitor
Execution Engine
51
CQSA Discussion

• Initial approach
• Prediction technique should be chosen
• Define language describing adaptation policies
  corresponding actions
• Extend approach to SBAs/composite services

52
References (I)

• Degwekar et. al. 2004
• S. Degwekar, S. Y. W. Su, and H. Lam, Constraint
  specification and processing in web services
  publication and discovery, in ICWS 04. San
  Diego, CA, USA IEEE Computer Society, 2004, pp.
  210217.
• Cortes et. al. 2005
• A. R. Cortes, O. Martin-Diaz, A. D. Toro, and
  M. Toro, Improving the automatic procurement of
  web services using constraint programming, Int.
  J. Cooperative Inf. Syst., vol. 14, no. 4, pp.
  439468, 2005.

53
References (II)

• 1
• M. Comuzzi, K. Kritikos, and P. Plebani, A
  semantic based framework for supporting
  negotiation in Service Oriented Architectures,
  in IEEE CEC 2009. Vienna, Austria IEEE Computer
  Society, 2009.
• 2
• K. Kritikos and D. Plexousakis, Evaluation of
  qos-based web service discovery algorithms, in
  IEEE Transactions on Services Computing, 2009.

54
References (III)

• 3
• K. Kritikos, QoS-based web service description
  and discovery, Computer Science Department,
  University of Crete, Heraklion, Greece, PhD
  Thesis, December 2008.
• 4
• C. Muller, A. Ruiz-Cortes, and Manuel Resinas,An
  Initial Approach to Explaining SLA
  Inconsistencies, in ICSOC 2008. Sydney,
  Australia Springer, 2008.