ITR: Dynamic Code Enhancement and Scheduling Techniques for Complex Simulations

当今最重要的一类程序是大规模、自适应和时间相关的模拟。对于解决重要的科学问题，如粒子动力学和边界元问题，这些问题变得越来越重要。运行任何代码的另一个同样重要的环境是工作站集群。这样的系统可以包括支持混合消息传递/共享地址空间编程范例的单处理器和SMP节点。当复杂的代码在如此复杂的环境中运行时，该项目将提供自动化和半自动的工具来优化串行性能、并行性能和整体资源利用率。从技术上讲，该项目的目标是开发一个全面的动态代码增强、资源管理、调度和性能监控框架。这是通过将许多代码优化和调度决策降级到运行时来实现的，在运行时，它们可以依赖于性能跟踪。该框架将进程-线程模型推广为可调度实体模型，其中进程和线程分别被视为自由实体和绑定实体。程序员或编译器都可以创建这些实体。此外，动态代码增强器在运行时对线程执行粒度控制的优化，并在自由实体和绑定实体之间进行转换以提高性能。动态代码增强的开销在几个计算步骤中分摊。聚合调度器/资源管理器将指定的实体映射到主机，使用性能数据来优化其对CPU、存储器系统、网络和并行代码性能的决策。动态代码增强器/调度器框架由自动插入到代码中的在线性能监控触发。

One of the most important classes of programs today is large-scale adaptive and time-dependent simulations. These are increasingly important for solving important scientific problems such as particle dynamics and boundary element problems. An equally important environment for running any code is a cluster of workstations. Such a system may include single-processor and SMP nodes supporting a hybrid message passing/shared address space programming paradigm. This project will provide automated and semi-automated tools for optimizing serial performance, parallel performance, and overall resource utilization when those complex codes are run in such complex environments.Technically, the goal of the project is to develop a comprehensive dynamic code enhancement, resource management, scheduling, and performance monitoring framework. This is accomplished by relegating a number of code optimization and scheduling decisions to run-time, where they can rely on performance traces. The framework generalizes the process-thread model to a schedulable entity model in which processes and threads are treated as free and bound entities respectively. Either the programmer or the compiler can create these entities. In addition, the dynamic code enhancer performs optimizations at run-time granularity control for threads, and transformations between free and bound entities to improve performance. The overhead of dynamic code enhancement is amortized over several computation steps. An aggregate scheduler/resource manager maps the specified entities to hosts, using performance data to optimize its decisions for CPU, memory system, network, and parallel code performance. The dynamic code enhancer/scheduler framework is triggered by online performance monitoring that is automatically instrumented into the code.