Taking advantage of usage service level agreements for better Grid resource scheduling.

摘要

In this dissertation I propose, design, build, and evaluate a different approach for controlled resources sharing in large distributed systems based on usage service level agreements (uSLAs). The proposed model is targeted for large and dynamic distributed environments and is itself distributed in order to cope with large communities of users that reside in different administrative domains. Agreements and Service level Agreements are not new concepts at this time. However, applying these concepts in a large scale Grid that makes scheduling and resource sharing effective is difficult. Without such an agreement based resource sharing mechanism in place, existing Grid scheduling, centralized or distributed, either do not scale well, or are not effective due to a high overhead of gathering up-to-date resource availability information. My thesis is that the explicit representation, enforcement, and management of uSLAs can serve as an objective organizing principle for such systems. uSLAs express how resources must be used over time intervals and represent a novelty for the Grids. The concept comes from the networking domain, where bandwidth is allocated based on specific rules.;The main objective is resources neither to remain idle when there are available workloads for execution, nor a virtual organization (VO) to consume more computing resources than provided. In support of this thesis, my contributions are as follows. First, I propose new mechanisms for uSLA specification and enforcement at various levels and perform experimental measurements of the improvements that these mechanisms can enable in different environments and for different workloads. Second, based on the real deployments, I introduce a method for determining uSLAs via observation rather than specification, that is, an algorithm that a client can use to determine automatically the uSLA that is delivered by a resource in practice. Third, I introduce GangSim, a simulator for Grid scheduling studies that allows uSLAs to be specified and simulated at different levels, as well as automated performance measurements. And fourth, I present GRUBER, a Grid resource scheduling prototype and architecture that allows uSLAs to be specified by site, VO, and, group administrators.