CSR: Small: Adaptive Synchronization for Multicore Systems

事务性内存（TM）是简化正确并行程序开发的最有前途的技术之一。已经开发了几十种不同的技术来实现TM，每种技术似乎都非常适合应用程序和硬件特性的某种组合。不幸的是，当为工作负载选择错误的算法时，可能会导致异常糟糕的性能。在行为随时间变化的程序中，这个问题尤其严重：一个执行阶段的最佳TM实现可能在另一个阶段不可接受。本研究将创建新的动态自适应算法，使TM的实现可以在程序执行过程中反复更改，从而在每个程序阶段实现性能最大化。研究将从机制和政策两个方面进行探讨。它将为选择TM实现创建新的算法和启发式方法。它还将对多种体系结构上的各种TM算法和工作负载的行为进行建模，以便创建适合指导自适应的知识库。为了实现最大的性能，本研究将考虑一系列广泛的特征，包括TM实现细节、瓶颈和开销；病理风险；硬件环境的属性；以及工作负载需求，例如线程间同步、I/O和对共享数据的非事务性访问的频率。本研究还将开发新的静态分析和动态优化框架，以避免任何开销，同时提供健壮的自适应TM。原型和源代码将作为开源软件分发。

Transactional Memory (TM) is among the most promising techniques for simplifying the development of correct parallel programs. Dozens of different techniques have been developed to implement TM, with each appearing ideally suited to some combination of application and hardware characteristics. Unfortunately, when the wrong algorithm is chosen for a workload, pathologically bad performance can result. The problem is particularly acute in programs whose behavior varies overtime: the best TM implementation for one phase of execution may be unacceptable during another phase. This research will create new algorithms for dynamic adaptivity, so that the implementation of TM can be changed repeatedly during program execution, thereby maximizing performance during each program phase.The research will explore both mechanism and policy. It will create new algorithms and heuristics for selecting TM implementations. It will also model the behavior of a wide array of TM algorithms and workloads on multiple architectures, in order to create a knowledge base suitable for guiding adaptivity. To achieve maximum performance, the research will consider a broad set of characteristics, to include TM implementation details, bottlenecks, and overheads; risk of pathology; properties of the hardware environment; and workload requirements such as frequency of inter-thread synchronization, I/O, and nontransactional access to shared data. This research will also develop novel static analyses and a dynamic optimization framework to avoid any overhead while providing robust, adaptive TM. Prototypes and source code will be distributed as open-source software.