Scalable Concurrent Compacting Garbage Collection for Commodity Multi-Core Processors

随着商用并行多核处理器（如Intel的Core处理器）进入主流，并发性是一个日益重要的研究领域。同时，Java和c#等主流编程语言既支持并发编程，也支持自动内存管理（所谓的“垃圾收集”或GC）。普通程序员期望他们的编程平台能够为运行在商用并行计算机上的应用程序提供响应性和性能。这个项目探索了GC的新方法，这些方法利用硬件并行性来保留和提高程序所需的空间/时间性能特征。它将通过开发新的多核产品使硬件行业受益，同时通过提高软件的可用性、性能和可靠性为经济和国家安全做出贡献。该项目将培养高需求的毕业生，他们精通并发编程的艺术和科学，能够在从桌面到数据中心的更可靠的高性能应用系统的生产中利用并行性，从而扩大国家技术劳动力的能力。虽然高效的并发非移动收集器在生产编程语言中无处不在，但高效的非侵入式压缩仍然缺失。没有压缩（无论是通过GC还是单独压缩），最大化内存利用率和局部性以及最小化内存碎片都是困难的。保持响应性和利用并行性意味着协调压缩与并发性。该项目正在设计自动压缩动态分配数据的新方法，重点是允许压缩与应用程序同时（即并行）发生的技术。基于并发复制（即移动）GC以及独立并发压缩（与非移动GC一起）的并发压缩方法正在为已建立的基准程序设计、实现和评估，这些程序使用度量响应性、应用程序利用率和吞吐量的指标。

Concurrency is an increasingly important area of study as commodity parallel multi-core processors (such as Intel's Core processors) enter the mainstream.At the same time, mainstream programming languages such as Java and C# support both concurrent programming and automatic memory management (so-called "garbage collection" or GC). Ordinary programmers expect their programming platforms to yield both responsiveness and performance for their applications running on commodity parallel computers. This project explores new approaches to GC that exploit hardware parallelism to retain and improve desirable space/time performance characteristics for programs. It will benefit the hardware industry by exploiting their new multi-core offerings, while contributing to economic and national security by enhancing usability, performance, and reliability of software. The project will produce high-demand graduates skilled in the art and science of concurrent programming, able to exploit parallelism in the production of more reliable high-performance application systems, from the desktop to the data center, so expanding the capabilities of the national technical work-force.While efficient concurrent non-moving collectors are ubiquitous in production programming languages, efficient non-intrusive compaction is still missing. Without compaction (whether by the GC or separately), maximizing memory utilization and locality, and minimizing memory fragmentation is difficult. Retaining responsiveness and exploiting parallelism, means reconciling compaction with concurrency. The project is devising new approaches for automatic compaction of dynamically allocated data, with emphasis on techniques that allow compaction to occur concurrently (i.e., in parallel) with the application. Concurrent compaction approaches based on concurrent copying (i.e., moving) GC as well as standalone concurrent compaction (in concert with non-moving GC) are being designed, implemented, and evaluated for established benchmark programs using metrics that measure responsiveness, application utilization, and throughput.