Career: Adaptive Parallel Processing for Dynamic Applications

现有的重构编译器已经成功地通过使用当前的相关性测试和符号方法来检测常规程序的并行性，即具有静态(编译时)定义的相对简单的访问模式的代码。不幸的是，对于超过50%的程序来说，访问模式要么非常复杂，要么在执行前完全没有定义。例如，SPICE、DYNA-3D、GASSUS、DMOL、CHARMM和FIDAP就是这种“消耗周期”的应用程序。随着建模技术变得更加复杂，未来的大规模模拟本质上将主要是动态的，从而增加了静态不可分析代码的比例，从而有效地减少了当前编译器技术的覆盖范围。通过经典的编译时分析技术，稀疏的动态程序从根本上是难以处理的，这使得并行化或其他优化变得不可能。这项工作将通过创建一个框架来将当前技术的覆盖范围扩展到动态程序，该框架将使用运行时分析和优化技术来补充静态编译。还将研究应用级可编程硬件对某些简单、重复性操作的加速。最后，在自适应策略中，静态和运行时方法将与硬件和软件性能监视器和统计数据库连接在一个自适应循环中，该策略将在执行期间将应用程序调优到其数据。

Existing restructuring compilers have been successful in detecting the parallelism of regular programs, i.e., codes with relatively simple access patterns that are statically (compile time) defined, by using current dependence tests and symbolic methods. Unfortunately, for more than 50% of programs the access pattern are either very complex or are completely undefined before execution. Examples of such "cycle consuming" applications are SPICE, DYNA-3D, GAUSSIAN, DMOL, CHARMM and FIDAP. As modeling techniques become more sophisticated, future large scale simulations will be mostly dynamic in nature, thus increasing the fraction of statically non-analyzable codes and therefore effectively decreasing the coverage of current compiler technology. Sparse, dynamic programs are fundamentally intractable through classic compile-time analysis techniques making parallelization or other optimizations impossible. This work will extend the coverage of current technology to dynamic programs by creating a framework that will complement static compilation with run--time analysis and optimization techniques. Application level programmable hardware for the acceleration of certain simple, repetitive operations will also be studied. Finally, an adaptive strategy in which static and run-time methods will be connected in an adaptive loop with hardware and software performance monitors and statistical databases will tune the application to its data during execution.