CSR: Small: Collaborative Research: Pursuing High Performance on Clouds and Other Dynamically Heterogeneous Computing Platforms

新兴技术催生了革命性的计算平台，如计算云和许多类型的计算网格。这些平台承诺让公众能够接触到高性能计算平台。然而，要实现这一承诺，需要解决平台的动态异构性，即其组成计算机的相对能力和速度可能在不可预测的时间和以不可预测的方式变化的事实；例如，共享计算机可能会因为工作负载的不可预测变化而显著减慢或加快。对个人信息系统的研究正在开发一种变革性的计算范式，使高性能计算能够在这些平台上实现。新的范式取代了传统的调度器试图适应计算平台的特殊性-这是动态异构性所困扰的目标-通过在Away中协调复杂的计算来尊重计算的内在结构的相关细节。通过这种方式，该范例增加了并行(即同时)执行独立任务的机会，从而更快地完成计算。初步评估-通过使用通用计算调度器的模拟比赛-表明新的范例完成计算的速度通常比竞争策略快两位数的百分比。挑战是以计算效率的方式实现这一范式，而私人投资机构正在寻求各种途径来实现这种效率。由于动态异构性平台的复杂性，这一目标需要调度算法和仿真软件技术的进步。由于其在算法、应用程序和系统方面的互补专业知识，PI是唯一有资格产生这种进步的机构。

Emerging technologies have led to revolutionary computing platformssuch as computing clouds and many genres of computing grids. Theseplatforms promise to make high-performance computing platformsaccessible to the public. Realizing this promise, though, requiresone to cope with the platforms' dynamic heterogeneity, i.e., the factthat their constituent computers' relative powers and speeds canchange at unpredictable times and in unpredictable ways; for instance,shared computers may slow down or speed up significantly because ofunpredictable changes in workloads. The research of the PIs isdeveloping a transformative computing paradigm that will enablehigh-performance computing on these platforms. The new paradigmreplaces traditional schedulers' attempts to accommodate theparticulars of a computing platform---a goal that dynamicheterogeneity confutes---by orchestrating a complex computation in away that honors the relevant details of the computation's inherentstructure. In this way, the paradigm increases opportunities forexecuting independent tasks in parallel (i.e., simultaneously),thereby completing computations faster. Preliminary assessments---viasimulated competitions with common computation schedulers---suggestthat the new paradigm often completes computations faster thancompeting strategies by double-digit percentages. The challenge is torealize the paradigm in a computationally efficient manner, and thePIs are pursuing a variety of avenues toward achieving suchefficiency. Because of the complexity of dynamically heterogeneousplatforms, this goal requires advances in the technology of bothscheduling algorithms and simulation software. The PIs are uniquelyqualified to produce such advances because of their complementaryexpertise in algorithms, applications, and systems.