CAREER: Timeliness as a Controllable Dimension via Knowledge-driven System Management

现代生活方式的渗透和很大程度上是由与计算系统的共生互动决定的。期望它们在正确的时间（及时性）采取有意义的行动（逻辑正确性）。对于安全至关重要的自主系统，如无人驾驶车辆和飞机、制造业和发电厂，以及重症监护和生命支持设备，及时性至关重要。类似地，企业和云系统中的服务质量主要包含应用程序工作负载的时间特征。不幸的是，过去十年中对智能系统的强烈推动导致了软件和硬件层复杂性的爆炸式增长。反过来，系统复杂性的急剧上升大大削弱了我们推断、预测和控制应用程序工作负载时效性的能力。这个教师早期职业发展（Career）项目旨在通过综合研究和教育计划来协调对时效性的迫切需求和现代智能系统日益增长的复杂性。研究的重点是将时效性转化为一个维度，在高性能平台中可以直接和明确地进行管理。它通过认识到收集有关软件和硬件组件之间低级相互作用的精确知识并采取行动的重要性来做到这一点。研究了收集、保留和利用这些知识的系统范例，并特别关注可以立即应用于商业平台的实用技术。作为补充，PI还设计了一些教育活动，将时效性作为计算机科学课程的基本设计时间维度，并为大学预科学生设计了一些拓展活动，以探索时间敏感系统中的挑战。通过一个协调的计划来重新思考硬件和平台软件组件的角色，PI将为知识驱动的系统管理制定蓝图，并以明确控制时效性为目标。包括传统处理元素和可重新编程逻辑的新兴商业现成平台为实际实例化上述蓝图提供了独特的机会。三个正交的研究重点构成了这项研究工作的支柱。第一个重点是设计和开发一组启用机制，以最小的开销观察和控制性能关键型硬件资源。特别是，观察机制的设计是为了收集微观尺度应用程序-平台交互的见解。相反，控制机制的设计是为了在微体系结构级别上执行细粒度的资源供应决策。第二个要点解决了从细粒度数据中提取知识的挑战。这需要将获取的数据流置于上下文中，以构建及时应用程序进度的概念。利用及时的进度信息来检测过去的供应不足/过剩，并相应地执行分配/回收。同时，探讨了准确预测资源分配策略对未来应用进程影响的策略。最后，第三个要点涉及多个应用程序工作负载的临界感知组合，这些工作负载具有指定的时效性需求、对硬件资源的概要需求以及响应资源供应变化的已知时间行为。这个项目的意义有三个方面。首先，通过知识驱动的系统管理直接控制复杂应用程序的时效性，为数据驱动的实时决策提供了必要的原则基础。其次，作为这项工作的一部分开发的方法和技术将作为未来硬件平台的参考，通过设计实现对时效性的控制。第三，针对波士顿大学预科生、本科生和研究生的跨领域活动旨在将时间行为分析和时效性带入计算机科学教育的前景。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The modern lifestyle is permeated and largely defined by a symbiotic interaction with computing systems. These are expected to take meaningful actions (logical correctness) at the right time (timeliness). Timeliness is crucial in safety-critical autonomous systems such as driverless vehicles and aircraft, manufacturing and power plants, and devices for intensive care and life support. Similarly, quality-of-service in enterprise and cloud systems primarily encompasses the temporal characteristics of application workloads. Unfortunately, the intense push for intelligent systems over the last decade has caused an explosion in the complexity of software and hardware layers. In turn, the sharp rise in system complexity has significantly weakened our ability to reason about, predict, and control the timeliness of application workloads. This Faculty Early Career Development (CAREER) project aims to reconcile the critical need for timeliness with the growing complexity of modern intelligent systems with an integrative research and education plan. The research focuses on turning timeliness into a dimension that can be directly and explicitly managed in high-performance platforms. It does so by recognizing the importance of gathering and acting upon precise knowledge about the low-level interplay between software and hardware components. Systemic paradigms to collect, retain, and leverage such knowledge are studied, with a specific focus on practical technology that can be immediately employed on commercial platforms. In a complementary way, the PI has devised educational activities to introduce timeliness as a fundamental design-time dimension in computer science curricula, with a number of outreach activities designed for pre-college students to explore challenges in time-sensitive systems.Through a coordinated plan to rethink the role of hardware and platform-software components, the PI will lay down the blueprints for knowledge-driven system management with explicit control over timeliness as a goal. Emerging commercial-off-the-shelf platforms that include traditional processing elements and reprogrammable logic provide a unique opportunity to practically instantiate said blueprints. Three orthogonal research thrusts comprise the pillars of this research effort. The first thrust focuses on designing and developing a set of enabling mechanisms to observe and control performance-critical hardware resources with minimal overhead. In particular, observation mechanisms are designed to collect insights into micro-scale application-platform interactions. Conversely, control mechanisms are devised to enforce fine-grained resource provisioning decisions at a micro-architectural level. The second thrust tackles the challenge of extracting knowledge from fine-grained data. This entails contextualizing the acquired data streams to construct a notion of timely application progress. Timely progress information is leveraged to detect past under-/over-provisioning and perform allocation/reclaiming accordingly. Simultaneously, strategies to accurately predict the impact of resource provisioning policies over future application progress are explored. Finally, the third thrust concerns the criticality-aware composition of multiple application workloads with specified timeliness requirements, profiled demand for hardware resources, and known temporal behavior in response to changes in resource provisioning. The significance of this project is threefold. First, direct control over the timeliness of complex applications via knowledge-driven system management provides the necessary principled foundations for data-driven real-time decision-making. Second, the methodologies and techniques developed as part of this effort will serve as a reference for future hardware platforms to achieve control over timeliness by design. Third, cross-cutting activities targeting pre-college, undergraduate, and graduate students at Boston University aim to bring temporal behavior analysis and timeliness into the foreground of computer science education.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.