a probabilistic quality of service model for nondeterministic … · 2020. 6. 20. ·...
Post on 24-Feb-2021
6 Views
Preview:
TRANSCRIPT
A probabilistic quality of service model for nondeterministic service compositions
Doctoral thesis defense
Candidate: Adrian Satja Kurdija
Faculty of Electrical Engineering and Computing
Supervisor: Assistant professor Marin Šilić, PhD
June 2020
Outline
• Service-Oriented Architecture
• Quality of Service (QoS)
• QoS prediction
• Service selection
• Compositional QoS model
• Multi-Criteria Service Selection for Multi-User Composite Applications
• Evaluation
• Conclusion
2/50
Introduction
• Cloud computing applications contain various service invocations
• Example: an application such as Gmail uses various services (mail, chat/hangouts, calendar, tasks, translation)
• Multiple atomic services with different functionalities
• Application can be seen as a service composition
3/50
Service-Oriented Architecture
4/50
Fig. Multi-tenant travel booking service-based system (SBS) [1]
Service-Oriented Architecture
5/50
Fig. Multi-tenant travel booking service-based system (SBS) [1]
Service-Oriented Architecture
6/50
Fig. Multi-tenant travel booking service-based system (SBS) [1]
Service-Oriented Architecture
• Service Oriented Architecture (SOA): architectural style that assumes a variety of atomic reusable services which provide certain functionality through their publicly accessible interfaces
• Advanced functionality is achieved through atomic services composition into more complex composite services
7/50
Quality of Service (QoS)
• Each functionality → many service instances on servers worldwide → service class
• Functionally equivalent service candidates can have different non-functional properties, referred in literature as Quality of Service (QoS)
• QoS properties include:– availability
– reliability
– response time
– throughput
– reputation
– cost
8/50
Quality of Service (QoS)
• QoS properties depend on:– user-specific parameters
• location, network and device capabilities, usage profiles
– service-specific parameters
• location, computational complexity, system resources
– environment-specific parameters
• service provider load, network performance
• User requirements for QoS values:– maximal response time (e.g. ≤ 1 second)
– minimal reliability (e.g. ≥ 98%)
– minimal throughput (e.g. ≥ 1.5 requests per second)
– etc.
9/50
Service Composition
• Workflow of the composition - execution plan of tasks to perform
• Contains compositional structures:– sequence, branching, parallel execution, loop
• QoS of the composition depends on:– the actual execution path (probabilistic in case of e.g. branching)
– QoS of the selected service instances for each task
10/50
Fig. Example of a service composition (execution plan)
Research problem
• Which service instance to invoke at a particular time for a particular user?
• Two independent research tasks:– QoS prediction
• for atomic service instances
– Service selection
• for composite applications
11/50
QoS prediction
• Happens before the service selection as its prerequisite
• Most QoS(user, service) values are unknown (no data)
• Predictions are based on past invocation data for similar users and/or services (collaborative filtering)
12/50
QoS User 1 User 2 User 3 User 4
Service 1 ? ? 0.98 0.993
Service 2 ? 0.9 ? ?
Service 3 ? ? 0.99 ?
Service 4 0.97 ? 0.95 ?
Service 5 ? 0.96 ? 0.94
Service 6 0.992 ? ? ?
Fig. User-Service QoS data
QoS prediction
13/50
Fig. QoS Prediction (high-level)
QoS prediction model
• From Kurdija, Silic, Delac, Srbljic (2018): Real-time adaptive QoS prediction using approximate matrix multiplication [2]– Faster than standard UPCC/IPCC/Hybrid approaches with comparable
accuracy
14/50
Fig. QoS prediction model overview
Service selection problem
• Which service instance to select for a particular user and a particular task?
15/50
Fig. Multi-user service selection
Service selection problem
• Which service instance to select for a particular user and a particular task?
16/50
Fig. Multi-user service selection
Service selection problem
• Which service instance to select for a particular user and a particular task?
17/50
Fig. Multi-user service selection
Service selection constraints
• Optimize multiple application instances (users)
• Respect the throughput limit of each service instance– maximum number of requests it can process in a given time frame
• Each user might have constraints on QoS– total reliability / price / response time (…)
• A service instance might have good reliability, but high price, etc. → multicriteria service selection
18/50
Existing solutions
• Greedy approach: select a service with best QoS for each user?
• Problems:– QoS has many properties → possible trade-offs (e.g. price vs.
reliability)
– what matters is QoS of the whole composition, not just of a single task
– composition is non-deterministic
– service has a throughput limit → greedy will overload the most “popular” services
19/50
Existing solutions
• Reduction to Mixed Integer Programming (MIP) [4]– heuristic enhancements in case of large search space:
• heuristic ranking of services
• service clustering
• Reduction to Assignment Problem (AP) [5]– for single service class, 1 request per user and service
– solved by Hungarian (Kuhn-Munkres) algorithm
20/50
QoS User 1 User 2 User 3 User 4
Service 1 X
Service 2
Service 3 X
Service 3 X
Service 3
Service 3 X
Fig. Assignment problem
Processing capacity
(THR = 4)
Limitations of existing solutions
• Efficiency problems:– exponentially large search space (in case of MIP)
– large problem size for high throughputs (in case of AP)
• Simplistic treatment of QoS compositions– ignoring branching and/or probabilities
– aggregated QoS is calculated by simple sum/product or by taking worst-case execution path
• Non-personalized QoS– assumption that QoS values depend only on a service and not on a
user
21/50
Limitations of existing solutions
22/50
multi-user
multi-task (service
composition)
user-dependent
QoS
main approach
worst-case complexity
He et al.(2015)
✓ ✓ × mixed integerprogramming
(MIP)
exponential
Alrifai et al.(2015)
× ✓ × enhanced MIP exponential
He et al. (2015)
✓ ✓ × greedy + MIP exponential
Wang et al. (2015)
✓ ✓ × clustering + MIP
exponential
Wang and Cheng(2015)
✓ × × clustering + MIP
exponential
Wang et al.(2015)
✓ × ✓ assignment problem
polynomial
our work (2019)
✓ ✓ ✓ transportationproblem
polynomial
The proposed model
• From Kurdija et al. (2019): Fast Multi-Criteria Service Selection for Multi-User Composite Applications [6]
• Main aims: generality, time efficiency
• Support service composition with:– a large number of service-based applications
– for a large number of different users
– with the aim of meeting all (or as many as possible) QoS requirements
– respect the physical limitation of maximum throughput for each service in the cloud
– satisfy the assumption of QoS dependence on a particular user
– take into account the probabilistic aspect of a composite application (compositional QoS model)
23/50
Compositional QoS model
• Assume users with potentially different nondeterministic composition plans and QoS requirements
• Consider all common compositional structures (sequence, conditional branching, parallelism, and loops)
• Consider estimated probabilities of branching and the expected number of loop repetitions
• Calculate the expected number of invocations of each service class
• Calculate the expected compositional QoS values for a particular selection of atomic services
24/50
Compositional QoS model
• 𝐸𝐼(𝐶) = vector of expected number of invocations for each service class in composition 𝐶
• 𝑄𝑜𝑆(𝐶) = vector of expected QoS values for a specific service selection
• Calculated recursively – example (branching):– If 𝐶 is a composition which branches into 𝐶1, 𝐶2, … , 𝐶𝑚 with respective
probabilities 𝑝1, 𝑝2, … , 𝑝𝑚, then:
𝐸𝐼 𝐶 =
𝑗=1
𝑚
𝑝𝑗 𝐸𝐼 𝐶𝑗 ,
𝑄𝑜𝑆 𝐶 =
𝑗=1
𝑚
𝑝𝑗 𝑄𝑜𝑆 𝐶𝑗 .
25/50
Reduction to transportation problem
• Transportation problem (abstract):– each supplier has a given number of items to ship
– each demander has a given number of items to receive
– each supplier-demander connection has a cost
– goal: find shipping distribution (how many items to ship for each connection) to minimize total cost
• In service selection:– requests are items
– services are suppliers
– users are demanders
• demand is the expected
number of requests
– QoS-based cost of connections
26/50
Reduction to transportation problem
• Algorithm for solving TP:– Find an initial (heuristic) solution using Vogel Approximation
Method (VAM)
– Iteratively improve the solution until it is optimal, using Transportation Simplex Method (TSM)
• In our context:– multiple service classes → multiple transportation problems
– how to define transportation cost?
– how to enforce global (compositional) QoS requirements?
27/50
Iterative heuristic algorithm
• For each service class (task), reduce the service selection problem to a transportation problem
• Non-locality: define utility cost in a transportation problem by taking into account other tasks in user’s composition
• Transportation problems for different tasks can be solved in parallel
• Update cost (put more "weight") for non-satisfied QoS properties and repeat the procedure
28/50
Iterative heuristic algorithm
29/50Fig. Selection example
Iterative heuristic algorithm
30/50Fig. Selection example
Iterative heuristic algorithm
31/50Fig. Selection example
Transportation problem utility cost
• How to define a global-aware cost of matching service 𝑖 to user 𝑢?
• Basic idea: If 𝑖 → 𝑢 is selected in task 𝑇𝑗, how difficult would it be to complete the selection?
• Namely, do we need low or high quality services in other tasks to satisfy the QoS requirement of user 𝑢?– Rank the services in each task by QoS
– If low-QoS services complete the requirement with 𝑖 → 𝑢, then 𝑖 → 𝑢 is
"easy" (low cost)
– If only high-QoS services complete the requirement with 𝑖 → 𝑢, then 𝑖 → 𝑢 is "difficult" (high cost)
32/50
Transportation problem utility cost
33/50
Fig. Illustration of the concept of matching difficulty 𝑖 → 𝑢for a fixed user 𝑢, service 𝑖 and QoS property 𝑘
Transportation problem utility cost
34/50
Fig. Illustration of the concept of matching difficulty 𝑖 → 𝑢for a fixed user 𝑢, service 𝑖 and QoS property 𝑘
Transportation problem utility cost
35/50
Fig. Illustration of the concept of matching difficulty 𝑖 → 𝑢for a fixed user 𝑢, service 𝑖 and QoS property 𝑘
High-level algorithm illustration (1/3)
36/50
High-level algorithm illustration (2/3)
37/50
High-level algorithm illustration (3/3)
38/50
Evaluation
• Generating a testing dataset– mixture of actual and artificial test data
• Testing two variations of the proposed approach:– Service Selection using Vogel Approximation Method (SS-VAM)
– Service Selection using Transportation Simplex Method (SS-TSM)
• Types of experiments:– Single-Task Experiment
– User-Independent-QoS experiment
– General experiment
39/50
Evaluation
• Comparing with existing approaches according to:– QoS requirement satisfaction:
• accuracy = avg. percentage of satisfied QoS reqs.
– obtained QoS values:
• QoS improvement = 𝑎𝑣𝑔.obtained QoS value− required QoS value
required QoS value
– execution time (efficiency)
• Current cloud computing efforts deal with an increasing amount of users and service instances → focus on efficiency (more than accuracy)
40/50
Single-Task Experiment
• Application = single service, no compositions
• Comparison with AP model (reduction to assignment problem) and MIP
• 500 users, 100 services with low (1-50) or high (50-1000) throughput limits
41/50
Single-Task Experiment
42/50
User-Independent-QoS Experiment
• Assume that QoSu,i = QoSv,i for different users u, v
• Testing against enhanced-MIP models which depend on this assumption (Clus2-MIP and Clus3-MIP)
• 100 users, 8 tasks, 300 services, varied QoS requirement difficulty
43/50
User-Independent-QoS Experiment
44/50
General Experiment
• Without special assumptions
• Testing against general MIP and its enhancement (Greedy MIP)
• 100 users, 8 tasks, 300 services, varied QoS requirement difficulty
45/50
General Experiment: QoS
46/50
General Experiment: time and scalability
47/50
Conclusions
• The proposed SS-TSM model is the dominating approach in most experiments because of a significant reduction of execution time
• The proposed SS-VAM approach can be faster in a single-task scenario with high service throughput limits
• The alternative AP approach (based on the reduction to assignment problem) can be faster in a single-task scenario when throughputs are lower
• In the general scenario, the proposed SS-TSM model shows best results
48/50
References
1. Q. He, J. Han, F. Chen, Y. Wang, R. Vasa, Y. Yang, and H. Jin, "Qos-aware service selection for customisable multi-tenant service-based systems: Maturity and approaches," in 2015 IEEE 8th International Conference on Cloud Computing, pp. 237-244, June 2015.
2. A. Kurdija, M. Silic, and S. Srbljic, “Real-time adaptive qos prediction using approximate matrix multiplication,” Int. J. Web Grid Serv., vol. 14, pp. 200-235, Jan. 2018.
3. H. Jin, H. Zou, F. Yang, R. Lin, and X. Zhao, "A hybrid service selection approach for multi-user requests," in 2012 IEEE 9th International Conference on Embedded Software and Systems, Liverpool, 2012, pp. 1142-1149.
4. Y. Wang, Q. He and Y. Yang, "QoS-Aware Service Recommendation for Multi-tenant SaaS on the Cloud," 2015 IEEE International Conference on Services Computing, New York, NY, 2015
5. Wang, S., Hsu, C., Liang, Z. et al. Multi-user web service selection based on multi-QoS prediction. Inf Syst Front 16, 143–152 (2014).
6. A. S. Kurdija, M. Silic, G. Delac and K. Vladimir, "Fast Multi-Criteria Service Selection for Multi-User Composite Applications," in IEEE Transactions on Services Computing. doi: 10.1109/TSC.2019.2925614
49/50The rest of references can be found in the doctoral thesis.
Thank you!
Questions?
50/50
top related