CSR: EDS: Small: Energy-Aware Redundancy Management

冗余是提高生产力和弹性的基本系统特性。每个这样的系统都有多个冗余配置选项，具有特定的资源需求和成本/质量权衡。该项目针对具有三种冗余类型的应用程序，即近似、复制和实现。应用程序被表示为相互依赖的服务的集合，这些服务在不同的资源可用性（特别是能源）下运行。每种服务可能以不同的成本提供不同质量的结果/答案，可能使用不同的硬件/软件资源。关键的见解是：(1)质量和成本需求必须在所有应用程序服务之间取得平衡，以及(2)用户需要能够指定和探索应用程序成本/质量权衡。该项目实施了一种新的编程架构，可以有效地管理三种冗余类型，以便在有限的能源/成本预算和可能的服务故障条件下产生最佳的应用结果。该项目的目标是电池驱动的智能手机和自主机器人系统，但也扩展到云、企业和高性能计算领域。应用程序被建模为双加权有向图，其中节点表示不同的服务及其质量级别，边缘表示这些服务及其级别之间的资源或输入/输出依赖关系。双权重对服务的能量/成本以及服务的重要性及其对用户的级别进行建模。服务选择问题的最优解决方案选择服务及其级别，使重要性度量在给定的能量/成本预算下最大化。使用0-1整数规划公式和最先进的整数规划工具来解决这个np完全问题。

Redundancy is a fundamental system characteristic that improves productivity and resilience. Each such system has multiple, redundant configuration options with their specific resource requirements and cost/quality tradeoffs. This project targets applications with three types of redundancies, namely approximation, replication, and implementation. Applications are expressed as collections of interdependent services that operate under varying resource availabilities, particularly energy. Each service may offer outcomes/answers of different quality at different costs, potentially using different hardware/software resources. The key insight is that (1) quality and cost requirements have to be balanced across all application services, and that (2) users need to be able to specify and explore application cost/quality tradeoffs. The project implements a new programming architecture that effectively manages the three types of redundancies to produce the best application outcomes under limited energy/cost budgets and possible service failure conditions. The project targets battery-operated smart phones and autonomous robotic systems, but extends to cloud, enterprise and high-performance computing domains. Applications are modeled as dual-weighted, directed graphs where nodes represent different services and their quality levels, and edges represent resource or input/output dependencies between these services and their levels. Dual weights model the energy/cost of a service and the importance of the service and its levels for the user. The optimal solution of the service selection problem picks services and their levels such that the importance metric is maximized under a given energy/cost budget. This NP-complete problem is solved using a 0-1 integer programming formulation and state-of-the-art integer programming tools.