Introduction

Idea of creating Blazar originated with two different use cases:

  • Compute host reservation (when user with admin privileges can reserve hardware resources that are dedicated to the sole use of a project)
  • Virtual machine (instance) reservation (when user may ask reservation service to provide him working VM not necessarily now, but also in the future)

Now these ideas have been transformed to more general view: with Blazar, user can request the resources of cloud environment to be provided (“leased”) to his project for specific amount of time, immediately or in the future.

Both virtual (Instances, Volumes, Networks) and hardware (full hosts with specific characteristics of RAM, CPU, etc) resources can be allocated via “lease”.

In terms of benefits added, Resource Reservation Service will:

  • improve visibility of cloud resources consumption (current and planned for future);
  • enable cloud resource planning based on current and future demand from end users;
  • automate the processes of resource allocation and reclaiming;
  • provide energy efficiency for physical hosts (both compute and storage ones);
  • potentially provide leases as billable items for which customers can be charged a flat fee or a premium price depending on the amount of reserved cloud resources and their usage.

Glossary of terms

Reservation is an allocation of certain cloud resource (Nova instance, Cinder volume, compute host, etc.) to a particular project. Speaking about virtual reservations, we may have not only simple, solid ones (like already mentioned instances and volumes), but also complex ones - like Heat stacks and Savanna clusters. Reservation is characterized by status, resource type, identifier and lease it belongs to.

Lease is a negotiation agreement between the provider (Blazar, using OpenStack resources) and the consumer (user) where the former agrees to make some kind of resources (both virtual and physical) available to the latter, based on a set of lease terms presented by the consumer. Here lease may be described as a contract between user and reservation service about cloud resources to be provided right now or later. Technically speaking, lease is a group of reservations granted to a particular project upon request. Lease is characterized by start time, end time, set of individual reservations and associated events.

Event is simply something that may happen to a lease. In most simple case, event might describe lease start and lease end. Also it might be a notification to user (e.g. about soon lease expiration) and some extra actions.

Rationale

Blazar is created to:

  • manage cloud resources not only right now, but also in the future;
  • have dedicated resources for a certain amount of time;
  • prepare for the peak loads and perform capacity planning;
  • optimize energy consumption.

Lease types (concepts)

  • Immediate reservation. Resources are provisioned immediately (like VM boot or moving host to reserved user aggregate) or not at all. If request can be fulfilled, lease is created and success status is returned. Lease should be marked as active or to_be_started. Otherwise (if request resource cannot be provisioned right now) failure status for this request should be returned.
  • Reservation with retries. Mostly looks like previous variant, but in case of failure, user may want to have several (configurable number) retries to process lease creation action. In this case request will be processed till that will be possible to create lease but not more than set in configuration file number of times.
  • Best-effort reservation. Also might have place if lease creation request cannot be fulfilled immediately. Best-effort mechanism starts something like scavenger hunt trying to find resources for reservations. For compute hosts reservation that makes much sense, because in case there are instances belonging to other project on eligible hosts, and without them there will be possible to reserve these hosts, Blazar may start instances migration. This operation can be timely and fairly complex and so different strategies may be applied depending on heuristic factors such as the number, type and state of the instances to be migrated. Also Blazar should assert that there are at least enough potential candidates for the migration prior to starting the actual migration. If Blazar decides to start migration, it returns success state and marks lease as in_progress, otherwise - failure. If this ‘hunting’ ends successfully before configurable timeout has passed, lease should be marked as active, otherwise its status is set to timedout.
  • Delayed resource acquiring or scheduled reservation. In this reservation type, lease is created successfully if Blazar thinks there will be enough resources to process provisioning later (otherwise this request returns failure status). Lease is marked as inactive till all resources will be actually provisioned. That works pretty nice and predictable speaking about compute hosts reservation (because hosts as resources are got not from common cloud pool, but from admin defined pool). So it is possible for Blazar to predict these physical resources usage and use that information during lease creation. If we speak about virtual reservations, here situation is more complicated, because all resources are got from common cloud resources pool, and Blazar cannot guarantee there will be enough resources to provision them. In this failure case lease state will be marked as error with appropriate explanation.