[论文笔记]Global and Local QoS Guarantee in Web Service Selection (BPM 2005)

Time: 4 hours
本文作者是 Danilo Ardagna (意大利米兰理工大学, 04年博士毕业并留校读博后, 现在是Assistant Professor)
Research Interests
"My research work focus on the design, prototype and evaluation of self-management systems for Web service based IT architectures. In particular, my work aims on
    the design of optimization algorithms for Web services composition;
    resource management of autonomic computing infrastructures."

该作者发表的 部分论文列表:
Active Energy-Aware Management of Business-Process Based Applications. Position Paper. ServiceWave 2008 Proceedings, 2008
Model Driven QoS Analyses of Composed Web Services. ServiceWave 2008 Proceedings, 2008.
Adaptive Service Composition in Flexible Processes, IEEE Transactions on Software Engineering, 2007  (gs:39)
PAWS: a framework for processes with adaptive Web services. IEEE Software. 2007.
SLA Based Resource Allocation Policies in Autonomic Environments. Journal of Parallel and Distributed Computing. Elsevier, 2007.
Dynamic Web Service Composition with QoS Constraints. International Journal of Business Process Integration and Management (IJBPIM), 2006.
Global and Local QoS Guarantee in Web Service Selection, BPM 2005 (gs:41)

本文的几个概念(S2)
Execution path: 同Zeng Liangzhao文中概念, 同一个execution path中不包含alternative branch, 但是可以包含parallel sequences.
Sub path: Execution path的一部分, 不包含parallel sequences.
Execution Plan: 一个Execution plan有唯一对应的Execution path, 一个Execution path可以有n个Execution plan
Global plan: 针对组合中所有task

(S2) 对服务组合中存在多个Execution path时获取全局plan的处理及存在问题
Zeng04#4.2.2 如果一个WS被多个Execution path包含, Zeng提出了通过识别"hot path", 并将hot path上绑定的WS实例做为解.
关于hot path的选择, Zeng提出的一种办法是将执行最频繁的Execution Path做为hot path.
本文作者给出了一个实例说明Zeng的这种方法得到的全局解在某些情况下会违法全局约束(而事实上存在满足全局约束的解).

核心部分
(S4.3) 本节是全文核心部分, 讨论了将WSC(Web Service Composition)建模为MMKP问题, 并使用MILP来解决.
ILP方法大同小异, 无非就是定义一个objective function, 若干constraints, 没什么花头. 与Zeng04相比, 本文主要有2个改进:
(1) 目标函数是各个execution plan分数与执行频率相乘后的和
(2) 对每个subpath的execution time都进行约束, 而Zeng04中只考虑critical path.
本文考虑的约束条件包括:
(1) Assignment constraints
(2) Task duration constraints 有前后序的WS要满足时间约束
(3) Local constraints
(4) Execution time constraints  每一个subpath都要满足
(5) availability constraint  通过ln线性化.
(6) Price and reputation constraints 对于availaility和constraints这种demand constraints要转化成capacity constraints

对于标准的QoS+WSC问题, 使用ILP方法来解决, 目前的工作差不多已经做到极致了, 个人觉得很难有什么新的突破了. 与Zeng04相比, 本文的ILP建模更加准确完整, 但是效率方面可能会稍差. 本文使用了CPLEX工具来解决整数线性规划, 实验部分将文中方法与Zeng04的局部优化方法做比较, 不过只比较了effectiveness, 没有比较算法本身的performance.

重优化
在MAIS系统中, 服务选择优化与执行是交叉(interleaved)进行的, 即针对WS的动态特点, 需要进行重优化, 触发reoptimization的情况有以下3种(S4.0)
(1) periodically
(2) the end user changes the service channel
(3) a WS invocation fails
reoptimization的过程与optimization是一样的, 已经完成的WS不需要被考虑, 因此问题规模会有所简化.

将服务组合建模为DAG图的几个条件
(S4.1) 作者提到了可以将服务组合建模为DAG图的几个要求
(1) 对cycle进行unfolding处理
(2) 只存在structured cycles(BPEL定义的流程满足这个条件)
(3) 只考虑XLANG-style process specifications (对这条, 我不太清楚)

最后罗列一下 本文的一些假设
(S1) assume that service descriptions are stored and retrieved from enhanced UDDI registries(提供QoS信息)
(S2) assume that a composite service is characterized by a single initial task and a single end task (简化讨论, 影响不大)
(S2) assume that cycles are unfolded according to the maximum number of iterations(这个假设不太准确, 作者在future work里也提到要改进)
(S3) assume that a Web service implements a single operation(简化讨论, 影响不大)
(S3) under the hypothesis that the aggregated value for quality dimensions can be evaluated as sum/average/product/min/max ... (不影响主体)
(S4.1) exceptions are not considered in the optimization problem
(S4.2) availability聚合时假设不同WS间是独立的
(S5) assume the price was proportional to service reputation and availability and inverselly proportional to the queue length(实验部分)

你可能感兴趣的:(web Service)