CSR-AES: Optimizations for Optimistic Parallelization Systems

目前大多数通用处理器都有两到四个核心，预计处理器芯片上的核心数量每一年半就会翻一番。只有在利用粗粒度并行性的情况下，应用程序才能在这样的处理器上高效运行。然而，到目前为止，并行编程只在处理数组和关系等结构化数据的应用程序中成功；不幸的是，大多数通用应用程序处理非结构化数据，如图形和树。该项目专注于利用无序和有序工作列表在非规则应用程序中产生的一种特殊类型的数据并行性。乐观并行化是在这类应用中获得并行性的关键机制。运行时系统用于管理乐观并行性，编译器分析用于优化并行执行。最后，在程序执行时使用动态代码专门化进一步优化程序。为了研究这种方法的可扩展性，该项目使用了基于现场可编程门阵列的多核硬件原型。如果成功，该项目将大大有助于解决为多核处理器编写软件这一紧迫问题。

Most current general-purpose processors have two to four cores, and the number of cores on processor chips is expected to double every one and half years. Applications run efficiently on such processors only if they exploit coarse-grain parallelism. However, parallel programminghas so far been successful only for applications that deal with structured datasuch as arrays and relations; unfortunately, most general-purpose applications deal with unstructured data such as graphs and trees. This project is focusedon the exploitation of a particular kind of data parallelism in irregularapplications that arises from the use of unordered and ordered worklists. Optimistic parallelization is the key mechanism for obtaining parallelism in such applications. A runtime system is used to manage the optimistic parallelism, and compiler analyses are used to optimize parallel execution. Finally, programs are further optimized using dynamic code specialization as the program executes. To investigate the scalability of this approach, the project uses multicore hardware prototypes based on field-programmable gate-arrays. If successful, the project will go a long way towards solving the pressing problem of writing software for multicore processors.