CAREER:Exploiting Binary Rewriting to Analyze and Alleviate Memory Bottlenecks for Scientific Applications

如今，共享内存多处理器(SMPS)的高性能集群被用来处理科学应用程序的大数据集。在这些SMPS上，结合消息传递和共享存储的混合编程模型通常比纯消息传递的效率低，尽管前者更适合于SMP体系结构。其创新之处在于依赖于动态二进制重写，即在应用程序执行时对其进行性能分析和调优。技术挑战是1)开发动态二进制重写框架，2)确定一致性流量导致在存在并行的情况下不确定的执行顺序，3)确定内存瓶颈，4)确定二进制文件热点数据引用之间的数据依赖关系，5)研究执行二进制程序的内存改进程序转换的可能性，6)评估大规模基准优化的优点等。关键的智力优势是为长期运行的应用程序提供额外的动态优化。这项工作更广泛的影响在于，它通过充分利用软件优化来应对处理器和主内存速度之间日益扩大的差距。

Today, high-performance clusters of shared-memory multiprocessors (SMPs)are employed to cope with large data sets for scientificapplications. On these SMPs, hybrid programming models combingmessage passing and shared memory are often less efficient than puremessage passing although the former fits SMP architectures moreclosely.The objective of this work is to determine the sources ofinefficiencies in utilizing memory hierarchies of SMPs and to optimizememory behavior. The novelty lies in the reliance ondynamic binary rewriting, i.e., performance analysis and tuning areperformed on the application while it executes.The technical challenges are to1) develop a framework for dynamic binary rewriting,2) determine coherence traffic resulting in misses fornon-deterministic orders of executions in the presence of parallelism,3) identify memory bottlenecks,4) determine data dependencies between data references forhot-spots for binaries,5) study the potential for memory-improving program transformationson the executing binary and6) evaluate the merits of optimizations for large-scale benchmarksamong others.The key intellectual merit is in providing additional, dynamicoptimizations for long-running applications. The broader impact ofthis work lies in its contribution to counter the increasing gapbetween processor and main memory speeds by fullyexploiting software optimizations.