XPS:FULL: MSLO-JS: Towards Multi-SLO-Guaranteed per-Job Scheduling in Datacenters

今天的数据中心支持具有不同服务级别目标(SLO)的各种应用程序，例如面向用户的应用程序，如网络搜索和灾难恢复，这些应用程序需要实时或接近实时的响应，需要严格的作业延迟和吞吐量保证。然而，由于缺乏适当的抽象，现有的SLO感知的作业资源提供方法是平台相关的，并且在设计上是反复试验的。建议的解决方案提供了一种新的抽象，它提供了一种有效的关注点分离，从而使开发独立于平台的、可移植的算法成为可能，这些算法将不同的作业SLO转换为准确的性能目标(也称为。预算)，从而实现SLO保证的作业资源供应。拟议的方法可以通过设计得到证明，预计将提供(A)具有高性能保证和成本效益的计算机系统和体系结构设计的重要见解；(B)在性能保证和资源利用方面的显著改进，并为日益流行的云计算环境降低运营成本。这项研究还有望鼓励学术界和产业界、跨学科和跨层次的合作。拟议的研究为实现SLO保证的就业资源提供奠定了坚实的理论基础。它探讨了基本的设计原则，按设计是跨层的，涉及从应用程序到运行时系统和系统体系结构的两层设计。在上层的应用层，任何给定的作业工作流以有向无环图(DAG)的形式表示，作业SLO被转换成DAG中各个任务节点的精确延迟预算，而与用于运行作业的底层系统无关。在较低的运行时系统层，选择各个任务节点的子系统，并分配资源以满足所有任务性能预算，从而满足作业SLO。这项拟议的研究将使我们能够开发具有SLO保证的作业资源提供算法，同时允许服务整合和实现高数据中心资源利用率。

Today's datacenters enable a wide range of applications with diverse service-level objectives (SLOs), e.g., user-facing applications such as web searches and disaster recoveries that require real-time or near-real-time responses, calling for stringent job latency and throughput guarantees. However, due to the lack of a proper abstraction, the existing SLO-aware job resource provisioning approaches are platform dependent and trial-and-error by design. The proposed solution presents a new abstraction that provides an effective separation of concerns and thus makes it possible to develop platform-independent, portable algorithms that translate diverse job SLOs into exact performance objectives (a.k.a. budgets) for constituent tasks, resulting in SLO-guaranteed job resource provisioning. The approach proposed can be proven by design and is expected to provide (a) important insights into computer systems and architecture designs with high performance guarantees and cost-effectiveness; and (b) significant improvements in performance guarantees and resource utilization, and the reduction of operating cost for the increasingly popular cloud computing environment. It is also expected to encourage academia-and-industry,interdisciplinary and cross-layer collaborations.The proposed research develops a sound theoretical foundation to enable SLO-guaranteed job resource provisioning. It explores fundamental design principles and is cross-layer by design, involving a two-layer design, from applications to runtime system and system architecture. At the upper, application layer, with any given job workflow represented in the form of Directed Acyclic Graphs (DAGs), the job SLOs are translated into precise latency budgets for individual task nodes in the DAG, independent of the underlying system to be used to run the job. At the lower, runtime system layer, the subsystems for individual task nodes are selected and the resources are allocated to meet all the task performance budgets and hence the job SLOs. This proposed research will enable us to develop job resource provisioning algorithms with SLO guarantee, while allowing for service consolidation and achieving high datacenter resource utilization.