VirTherm-3D - Formalization of Multi-Agent System Management and Adaptive Modular Redundancy for Dependable 3D MPSoCs

VirtTherm-3D是spp1500一期和二期项目的延续，名称相同。第一阶段侧重于基于I/O虚拟化的3D多核平台上的基本热管理，第二阶段则推广了基于代理的热系统管理，并实现了任意对任意通信虚拟化。第三阶段的重点是代理系统的形式化，以及阶段1和阶段2中3D平台的自适应模块化冗余。更具体地说，我们计划形式化代理系统，以便能够提供代理系统热管理行为的保证或概率，以确保3D多核系统在其生命周期内相对于短期和长期约束的可靠运行。特别是，我们关心的是一个平衡的管理，对短期热管理决策（即秒）及其对长期老化效应（即老化起作用的月/年）的影响敏感，这将需要正规化代理系统。因此，在不牺牲基于代理的系统管理的可伸缩性和灵活性的情况下提供高水平的可预测性是一个主要挑战。此外，我们引入了新的自适应模块化冗余（AMR）概念作为进一步提高可靠性的手段。AMR采用混合硬件/软件多层方法，通过3D多核平台上的代理系统隐式复制任务，具有最小的侵入性。因此，AMR允许在运行时根据系统的热、老化、功率和资源相关预算，通过动态软件任务实例化来模拟双重和三重模块化冗余保护。一个主要的研究问题将是确定实例化通用检查单元的正确数量和位置，以及处理同步和任务运行时变化。总之，第三阶段的VirTherm-3D建立在前两阶段的平台之上，进一步增强了多层可靠性覆盖范围，以及在实际嵌入式系统中部署所需的行为可预测性。

VirtTherm-3D is the continuation of the project with the same name in the SPP 1500 phases I and II. Whereas the first phase focused on basic thermal management on a 3D multi-core platform with I/O virtualization, the second phase generalized agent-based thermal system management and enabled any-to-any communication virtualization. The focus of this third phase is on a formalization of the agent system and on adaptive modular redundancy of the 3D platform from phases I and II. More specifically, we plan to formalize the agent system in order to be able to either give guarantees or probabilities of the agent system's thermal management behavior to ensure a reliable operation of the 3D multi-core system during its lifetime with respect to short and long term constraints. In particular, we are concerned about a well-balanced management sensitized to short-term thermal management decisions (i.e. seconds) and their impact on longer-term aging effects (i.e. months/years where aging plays a role) which will require formalizing the agent system. Thereby, it is a main challenge to provide a high level of predictability without sacrificing the scalability and flexibility of the agent-based system management. Furthermore, we introduce the new adaptive modular redundancy (AMR) concept as a means to further increase dependability. AMR follows a hybrid hardware/software multi-layer approach with minimal intrusiveness to implicitly replicate tasks through the agent system on the 3D multi-core platform. Thus, AMR allows emulating dual and triple modular redundancy protection through dynamic software task instantiation during runtime in compliance with the thermal, aging, power and resource related budgets of the system. A major research question will be to identify the right amount and locations to instantiate the generic checking units and coping with synchronization and task runtime variations. In summary, VirTherm-3D in its third phase builds upon the platform of the first two phases and further enhances multi-layer dependability coverage along with predictability in behavior necessary for deployment in real-world embedded systems.