Adaptive Software Substrate for High Performance Wide Area Computing

该项目的目标是为构建自适应运行时系统奠定基础，这些系统在运行时显示的各种资源条件（CPU，内存和网络）下表现良好。主要结果如下：（1）我们建立了一个有效的访问控制（例如，互斥）方法，它适应于应用程序中的并行度（发表在ACM PaCT上）。(2)我们提出了一个框架，在该框架中，并行应用程序可以从一组资源迁移到另一个在运行时，根据资源条件（已发布或将发布在ACM PPoPP和ACM/IEEE CCGrid）。(3)我们解决了资源选择问题，在该系统中，给定的计算/通信要求的应用程序和计算/通信能力的可用资源，试图自动选择良好的资源的应用程序。它提出了问题公式，一种假设资源需求和条件不随时间变化的算法，并进行了模拟。(4)我们建立了一个动态的内存管理方案，可以交易线程级的局部性分配的内存和总的内存需求。(5)我们为保守的垃圾收集器实现了非常短的暂停时间（小于10 ms），这在社区中被认为是困难的（发表在ACM ISMM上）。

The objective of this project was to establish foundations for building adaptive runtime systems, which behave well under various resource conditions (of CPU, memory, and network) that reveal at runtime. Main results are as follows, (1) We established an efficient access control (e.g., mutual exclusion) method for shared data, which adapts to the degree of parallelism in the application (published in ACM PaCT). (2) We proposed a framework in which parallel applications can migrate from a set of resources to another at runtime, depending on resource conditions (published or to be published in ACM PPoPP and ACM/IEEE CCGrid). (3) We tackled the resource selection problem in which the system, given computation/communication requirements of the application and computation/communication capacity of available resources, tries to select good resources for the application automatically. It proposed the problem formulation, an algorithm assuming resource requirements and conditions do not change over time, and conducted a simulation. (4) We established a dynamic memory management scheme that can trade thread-level locality of allocated memory and the total memory requirements. (5) We achieved a very short pause time (less than 10ms) for conservative garbage collectors, which have been considered to be difficult in the community (published in ACM ISMM).