A Probabilistic QoS Model and Computation Framework for Web Services-Based Workflows

1

• A Probabilistic QoS Model and Computation
  Framework for Web Services-Based Workflows
• San-Yih Hwang, H. Wang, J. Srivastava
• National Sun Yat-sen U., Taiwan
• Univ. of Minnesota, USA

2
Overview

• Introduction
• QoS metrics and modeling
• WS-Workflow QoS Computation
• Performance evaluation
• Conclusions

3
Web Service based workflows

• Web Service a modular and self-described
  application that uses Web technologies to
  interact with other services.
• Workflow a process by which a series of
  activities are executed in a specific sequence.
• WS-Workflow (composite web service) a workflow
  of activities, each of which is wrapped as a web
  service.
• e.g., a Travel Planner may aggregate multiple
  Web services for flight booking, travel
  insurance, accommodation booking, car rental,
  etc.
• Quality of Service (QoS) non-functional measures
  of a service, which often decides the
  satisfaction of a user toward the service .

4
Web service QoS Category Instance-level QoS Metrics Service Classes Metrics System-level QoS Metrics
1. Performance Response time time elapsed from the submission of a request to the time the response is received Throughput the number of instances completed per time unit
2. Resources Cost the amount of money paid for executing an instance Memory/CPU/ bandwidth
3. Dependability Reliability The probability that the service can be successfully completed. Availability The probability that a service can be successfully invoked. TTR time to repair
3. Dependability Reliability The probability that the service can be successfully completed. Availability The probability that a service can be successfully invoked.
4. Fidelity Reputation rating, usually measured as a scalar value (e.g., 1, 2, 3, 4, 5, the higher the better)
5. Transactional properties ACID properties
5. Transactional properties Commit protocol (e.g., 2PC)
6. Security Confidentiality
6. Security Nonrepudiation
6. Security Encryption
5
The problem

• The goal is to compute the QoS measures of a
  WS-workflow from those of its constituent web
  services.
• Four instance_level QoS metrics are considered
• Response time the time elapsed from the
  submission of a request to the time the response
  is received .
• Cost amount of money paid.
• Reliability the probability that the service can
  be successfully delivered.
• Fidelity reputation rating.
• How do we represent a QoS measure of a web
  service?
• A single value, used by most of the previous
  researches
• A probability distribution, adopted by us

6
The problem (Cont.)
W (parallel)

• Why not using a single value for a QoS measure of
  a web service?
• Instance-level QoS measures are inherently
  probabilistic.
• Choosing a single value for a QoS measure (e.g.,
  average case) may yield incorrect result.

A1
A2

• Response time of A1 N(10, 10)
• Response time of A2 N(10, 10)
• Average response time of W is NOT 10.

7
The problem (Cont.)
W (conditional)

• Why not using a normal distribution for modeling
  each QoS measure of a Web service?
• Some QoS measure may not follow normal
  distribution.
• Even if the QoS measures of all activities follow
  normal distribution, the QoS of a WS-workflow may
  not follow normal distribution, e.g., parallel,
  conditional selection

A1
0.5
0.5
A2

• Response time of A1 N(10, 5)
• Response time of A2 N(20, 5)

8
Probabilistic Modeling of WS QoS

• A QoS measure of a web service is modeled as a
  discrete random variable.
• Probability Mass Function (PMF)
• Let the sample space of X be Dom(X), then
• e.g. Suppose the probabilities of a web service w
  being completed in one, two, and three days, are
  0.2, 0.6, and 0.2, respectively. The PMF of its
  response time is
• fresponse_time(w)(1)0.2
• fresponse_time(w) (2)0.6
• fresponse_time(w) (3)0.2

9
WS-workflow QoS framework
WS invocation log
Web services

WS-workflow QoS Framework
WS selection
WS-workflow QoS Model
WS-workflow QoS Computation
WS-workflow enactment
WS SLA spec
WS-workflow QoS Objective Spec
WS-workflow QoS Monitoring
invokes
owner
user
10
Computing QoS Values of WS Compositions

• A structured workflow can be constructed
  recursively by the following 5 basic constructs.
• Sequential
• Parallel
• Conditional
• fault-tolerant
• Loop

11
Sequential

• Cost
• Time
• Reliability
• Fidelity

a1
a2
an

12
Parallel
a1, a2, , an are executed concurrently.

13
Conditional
Only one of the activities is executed.

,
14
fault-tolerant
All the activities are executed concurrently, but
only one of them need to be finished.

is the probability that ai is finished first.
15
Loop

• We model the number of iterations as a discrete
  random variable, e.g. 1 time with probability
  0.3, and 2 times with probability 0.7.
• It can be converted to a equivalent conditional
  structure.

16
Operations of discrete random variables

• Basic operations
• Sum
• Multiplication
• MAX
• MIN
• Conditional selection
• Let X, Y be independent random variables
• Dom(X)x1, x2, , xm , PMF fX(x)
• Dom(Y)y1, y2, , yn , PMF fY(y)

17
ZXY

• PMF

Dom(X)1, 2, fX(1)0.3, fX(2)0.7 Dom(Y)1
0, 20, fY(10)0.4, fY(20)0.6 Dom(Z)11, 12,
21, 22 fz(11)0.30.40.12, fz(12)0.70.40.28
fz(21)0.30.60.18, fz(12)0.70.60.42
Note Multiplication is similar, except that z is
the product of some x and y.
18
ZMax(X, Y)

• Dom(Z) Dom(X)?Dom(Y)

Note MIN operation is very similar to MAX.
19
Conditional selection

• exclusive
  selection of a random variable according to the
  associated probabilities.
• pi the probability that Xi is selected.

20
An example WS-workflow
21
Tree representation of a workflow

• A structured workflow can be represented by a
  tree.
• A composite activity can be substituted by an
  equivalent atomic activity.
• Use bottom-up method to compute the Workflow QoS.

22
Sample space reduction

• When combining the random variables, the sample
  space size of the resultant variable may increase
  dramatically.
• To keep the sample space at a reasonable size
  after some operation, we have to group the
  elements in the sample space. Each group is
  represented by one value.
• Find an optimal grouping scheme which minimizes
  the error.

23
The optimal solution

• Dynamic programming
• Let e(i, j, k) be the optimal aggregate error of
  partitioning xi, xi1, , xj into k subsequences.
  
• Recursive function
• 
  
• 
  if j-i1gtk and kgt1
• e(i, j, k) 0 if j-i1k
• e(i, j, 1) error(i, j).
• Time complexity O(s3m2), where s is the number
  of elements in the original domain and m is the
  number of elements in the domain after reduction.

24
Heuristic method

• Algorithm
• Find the pair of adjacent elements in the domain
  which has least error when merged.
• Replace the two elements by a new element.
• Repeat until the reasonable size is reached.
• Priority queue can be used to find the pair of
  samples with least error in O(logs) time.
• Time complexity O(slogs)

25
Performance evaluations

• Sample space size reduction
• Performance metric cumulative distribution
  function (CDF). The CDF of Z, denoted as FZ(z),
  is defined as Pr(Z?z)
• The following figure shows the difference of CDFs
  of each method (DP and greedy methods ) and the
  theoretical value when the size of PMF is
  reduced from 400 to 30.
• Mean error
• DP 0.001494, Greedy 0.002136

26
Response time accuracy

• Computation of response time of the experimental
  WS-workflow PC order fulfillment.

activity Mean Standard deviation minimum maximum Probability
HDD Proc. 1 2 1 1 4 0.8
HDD Proc. 2 3 1.5 1 5 0.2
CPU Proc. 2 0.8 1 4 N/A
CD-ROM Proc. 2 0.7 1 4 N/A
Assembly 2 0.5 1 3 N/A
Test 1 0.2 0.5 2 N/A
FixTest 0.5 0.2 0.1 1 N/A
Shipping 1 0.2 0.5 2 N/A
Email notification 0.6 0.2 0.2 1 N/A
Phone notification 0.5 0.2 0.1 1 N/A
27
Response time accuracy

• The following figure show the difference between
  the CDF of the greedy method and simulation
  result.
• Maximum error 0.008
• The greedy method is a thousand times faster than
  the simulation.

28
Cost accuracy
Activity Mean Standard deviation minimum maximum
HDD Proc. 1 100 6 80 120
HDD Proc. 2 110 10 80 130
CPU Proc. 150 5 130 160
CD-ROM Proc. 80 10 60 100
Assembly 20 3 10 30
Test 15 5 10 20
FixTest 10 3 5 20
Shipping 25 4 15 40
Email notification 1 0.5 3 5
Phone notification 2 1 1 3
29
Fidelity accuracy
1 2 3 4 5
HDD Proc. 1 0.006644 0.064938 0.259001 0.410416 0.259001
HDD Proc. 2 0.122595 0.233026 0.288758 0.233026 0.122595
CPU Proc. 0.027104 0.111310 0.259135 0.343315 0.259135
CD-ROM Proc. 0.153398 0.221476 0.250254 0.221476 0.153398
Assembly 0 0.001302 0.157605 0.683489 0.157605
Test 0 0.001302 0.157605 0.683489 0.157605
FixTest 0.006200 0.196167 0.595267 0.196167 0.006200
Shipping 0.022031 0.139256 0.349728 0.349728 0.139256
Email notification 0 0 0.369433 0.629267 0.001300
Phone notification 0 0 0.047833 0.904333 0.047833
30
Related work

• Estimating the QoS measures of a WS-workflow by
  assuming a single QoS value for each web service.
  Menace02 Cardoso02 Zeng03.
• Estimating the QoS measures of a WS-workflow
  using simulation Gillmann02Cardoso02
• Project management (CPM/PERT) for estimating
  response time

31
Summary

• We propose a probability-based WS-workflow QoS
  model and its computation framework.
• The computation framework is efficient and
  accurate.

32
Ongoing work

• Online QoS monitoring
• Online QoS estimation for a WS-workflow instance
• Pre-computation is needed for efficiency.
• Defining and ensuring QoS objective
• A QoS objective is a 5-tuple. E.g., (PC order
  fulfillment, response time, no larger, 7
  days, 90)
• Define a checkpoiont for each atomic web service.

33
Ongoing work

• Web service selection
• Each activity has a set of candidate web services
  that provide the same function but different QoS
  measures.
• Choose a web service for each activity such that
  some constraints are satisfied and the goal is
  optimized. Both constraints and the goal are
  specified in a probabilistic manner
• E.g., The probability that the entire WS-workflow
  can be completed in 10 days is at least 90.