Common use of Scheduling Model Clause in Contracts

Scheduling Model. Instance requests (to either start or stop an instance) are routed to a scheduler (as implemented by IaaS infrastructures such as Eucalyptus [6], Open Stack [9], and Cloud Stack [10]) which handles admission control and placement of instances on physical resources in a cluster of “nodes.” IaaS clouds typically define “instance types” that describe the resources that an instance will consume (CPU cores, memory, ephemeral disk storage, etc.). In the Eucalyptus systems (production and research) that we investigate in this work, we observe that the memory footprint associated with each instance type is such that the instance placement decision by the scheduler can be made strictly on core count. When an instance is admitted, the scheduler makes a place- ment decision by selecting a node on which the instance will run. In this study, we use simple first-fit placement in favor of more complex approaches to highlight the impact SLA-aware admission control. If an on-demand instance is requested, and the scheduler cannot find a node with available capacity, the scheduler selects one or more spot instances to terminate (evict) so that the on-demand instance can be scheduled. Further, our scheduler (like other Eucalyptus schedulers) assumes that the instance type definitions nest with respect to their core counts. For example, an empty 4-core node node is seen by the scheduler as having 1x 4-core slot, 2x 2-core slots, 4x 1-core slots, or 1x 2-core, 2x 1-core slots. The distinction between available cores and available slots is important when generating time-until-eviction estimates for different instance sizes and different cluster load levels.

Scheduling Model. Instance requests (to either start or stop an instance) are routed to a scheduler (as implemented by IaaS infrastructures such as Eucalyptus [6], Open Stack [9], and Cloud Stack [10]) which handles admission control and placement of instances on physical resources in a cluster of “nodes.” IaaS clouds typically define define “instance types” that describe the resources that an instance will consume (CPU cores, memory, ephemeral disk storage, etc.). In the Eucalyptus systems (production and research) that we investigate in this work, we observe that the memory footprint associated with each instance type is such that the instance placement decision by the scheduler can be made strictly on core count. When an instance is admitted, the scheduler makes a place- ment decision by selecting a node on which the instance will run. In this study, we use simple first-fit first-fit placement in favor of more complex approaches to highlight the impact SLA-aware admission control. If an on-demand instance is requested, and the scheduler cannot find find a node with available capacity, the scheduler selects one or more spot instances to terminate (evict) so that the on-demand instance can be scheduled. Further, our scheduler (like other Eucalyptus schedulers) assumes that the instance type definitions definitions nest with respect to their core counts. For example, an empty 4-core node node is seen by the scheduler as having 1x 4-core slot, 2x 2-core slots, 4x 1-core slots, or 1x 2-core, 2x 1-core slots. The distinction between available cores and available slots is important when generating time-until-eviction estimates for different instance sizes and different cluster load levels.