A distributed resource management architecture that supports advance reservations and co-allocation

TLDR

End‑to‑end QoS guarantees in emerging network‑based applications require mechanisms for dynamic discovery and advance or immediate reservation of heterogeneous resources that are independently controlled and administered. We propose the Globus Architecture for Reservation and Allocation (GARA) to address these challenges. GARA treats reservations and computational elements as first‑class entities, uniformly manages heterogeneous resources such as computers, networks, disk, and memory, and supplies application‑level co‑reservation and co‑allocation libraries that assemble resources dynamically according to QoS requirements and local policies. A prototype GARA implementation supports parallel computers, individual CPUs under a dynamic soft real‑time scheduler, and integrated services networks, and its performance results quantify the costs of the techniques.

Abstract

The realization of end-to-end quality of service (QoS) guarantees in emerging network-based applications requires mechanisms that support first dynamic discovery and then advance or immediate reservation of resources that will often be heterogeneous in type and implementation and independently controlled and administered. We propose the Globus Architecture for Reservation and Allocation (GARA) to address these four issues. GARA treats both reservations and computational elements such as processes, network flows, and memory blocks as first-class entities, allowing them to be created, monitored, and managed independently and uniformly. It simplifies management of heterogeneous resource types by defining uniform mechanisms for computers, networks, disk, memory, and other resources. Layering on these standard mechanisms, GARA enables the construction of application-level co-reservation and co-allocation libraries that applications can use to dynamically assemble collections of resources, guided by both application QoS requirements and the local administration policy of individual resources. We describe a prototype GARA implementation that supports three different resource type-parallel computers, individual CPU under control of the dynamic soft real-time scheduler, and integrated services networks, and provide performance results that quantify the costs of our techniques.