CAREER: Supporting Evolution in Distributed Object Computing Systems

分布式对象计算系统的编程由于对象持久存在并与系统中的许多其他对象交互的事实而变得复杂。当程序员更改和发展他们的对象时，系统其他部分的客户端都能感受到其影响，而不仅仅是在正在编程的应用程序中。因此，分布式对象程序员需要一种机制来管理这种演化，并将其告知系统中的其他对象。动态可配置分布式对象(DCDO)模型是在分布式对象系统中管理演化的第一步。在Legion网格计算系统中实现的DCDO模型有助于实现对象演化，并促进从多个独立实现组件开发分布式对象。使用DCDO，程序员可以演变现有的活动对象以接受新的成员函数，更改其成员函数的接口和行为，以及从其外部(公共)或内部(私有)接口中删除成员函数。程序员可以在不停用系统的任何部分、不替换二进制可执行文件、不中断演变对象的客户端、以及不必知道在对象最初被编译和运行时将发生什么改变的情况下，在运行中进行这些改变。该模型支持演化管理策略，这些策略定义对象类型何时以及如何从一个版本演化到下一个版本，并确定何时将类型更改传播到现有实例。构建在DCDO模型中的演化管理策略是管理变更的良好开端，但它们还不够。程序员需要一种更通用的机制来定义对象如何变化，以及限制演化如何发生。这项提案中描述的研究将解决这个问题。特别是，我们将使对象的演化特征成为分布式系统中的第一类实体，以便它可以被其他对象读取，然后这些对象可以根据它们的同行如何演化以及未来可能演化来适当地行为。客户端将能够读取服务器的演变历史，并能够确定对未来服务器演变施加的限制(如果有的话)。此外，客户端将能够向服务器注册回调功能，以便在发生更改时得到通知。当他们这样做时，客户端可以被赋予对何时以及如何执行更改的一些控制权。在某些情况下，客户端可能能够否决演进操作，或者将其延迟到客户端完成其当前应用程序。根据由在演进操作中注册了兴趣的客户端之间的回调功能发起的协议来实现该功能。这开始解决在构建为支持不同接口的客户端下更改实现的问题。在研究的另一个重要部分，我们将研究实际应用程序如何随着时间的推移而变化，以便有效地设计演化管理策略和约束。我们将在研究中包括广泛的不同应用程序，包括核心军团对象本身、高性能科学应用程序和分布式协作环境。在拟议计划的教育部分，我们将实施以学生为中心的通过行业项目进行应用学习(SCALIP)计划。在这个项目中，学生与当地行业合作，为其他学生在未来的核心研究生课程实例化中定义课堂项目。

Programming for distributed object computing systems is complicated by the fact that objects persist and interact with a number of other objects in the system. When programmers change and evolve their objects, the ramifications are felt by clients in other parts of the system, not just within the application that is being programmed. Therefore, distributed object programmers require mechanisms to manage this evolution and make it known to other objects in the system.The dynamically configurable distributed object (DCDO) model is a first step toward managing evolution in distributed object systems. The DCDO model, implemented within the Legion grid computing system, helps enable object evolution and facilitate the development of distributed objects from multiple independent implementation components. Using DCDOs, programmers can evolve existing active objects to accept new member functions, to change the interface and behavior of their member functions, and to remove member functions from their external (public) or internal (private) interface. Programmers can make these changes on the fly, without deactivating any part of the system, without replacing binary exe-cutables, without interrupting the clients of evolving objects, and without having to know what the changes will be at the time the objects are initially compiled and run. The model supports evolution management strategies that define when and how object types evolve from one version to the next, and determine when a type change is propagated to existing instances. The evolution management strategies built into the DCDO model are a good start toward managing change, but they are not sufficient. Programmers require a more general purpose mechanism for defining how objects can change, and for restricting how evolutions can take place. The research described in this proposal will address this problem. In particular, we will make the evolution characteristics of an object a first class entity in the distributed system, so that it can be read by other objects, which can then behave appropriately based on how their peers have evolved and may evolve in the future. Clients will be able to read the evolution histories of servers, and will be able to determine the restrictions, if any, that are placed on future server evolution. Further, clients will be able to register call-back functions with servers in order to be informed when changes take place. When they do, the clients can be given some control over when and how the changes can be carried out. In some cases, a client may be able to veto an evolution operation, or delay it until the client has finished its current application. This functionality is implemented in terms of a protocol that is initiated by the call back functions among the clients who have registered interest in the evolution operation. This begins to address the problem of implementations changing out from under clients that are built to support a different interface. In another important part of the research, we will study how real applications change over time, in order to effectively design evolution management strategies and constraints. We will include a wide range of different applications in the study, including the core Legion objects themselves, high performance scientific applications, and distributed collaborative environments. In the educational component of the proposed program, we will implement the Student Centered Applied Learning through Industry Projects (SCALIP) program. In this program, students work with local industries to define class projects for other students in future instantiations of core graduate courses.