Documentation ¶
Index ¶
- func PrintJobSummary(ctx *armadacontext.Context, prefix string, jctxs []*JobSchedulingContext)
- type GangInfo
- type GangSchedulingContext
- type GangSchedulingFit
- type JobSchedulingContext
- func (jctx *JobSchedulingContext) AddNodeSelector(key, value string)
- func (jctx *JobSchedulingContext) Fail(unschedulableReason string)
- func (jctx *JobSchedulingContext) GetAssignedNodeId() string
- func (jctx *JobSchedulingContext) IsSuccessful() bool
- func (jctx *JobSchedulingContext) SchedulingKey() (schedulerobjects.SchedulingKey, bool)
- func (jctx *JobSchedulingContext) SetAssignedNodeId(assignedNodeId string)
- func (jctx *JobSchedulingContext) String() string
- type PodSchedulingContext
- type QueueSchedulingContext
- func (qctx *QueueSchedulingContext) ClearJobSpecs()
- func (qctx *QueueSchedulingContext) GetAllocation() schedulerobjects.ResourceList
- func (qctx *QueueSchedulingContext) GetWeight() float64
- func (qctx *QueueSchedulingContext) ReportString(verbosity int32) string
- func (qctx *QueueSchedulingContext) String() string
- type SchedulingContext
- func (sctx *SchedulingContext) AddGangSchedulingContext(gctx *GangSchedulingContext) (bool, error)
- func (sctx *SchedulingContext) AddJobSchedulingContext(jctx *JobSchedulingContext) (bool, error)
- func (sctx *SchedulingContext) AddQueueSchedulingContext(queue string, weight float64, ...) error
- func (sctx *SchedulingContext) AllocatedByQueueAndPriority() map[string]schedulerobjects.QuantityByTAndResourceType[string]
- func (sctx *SchedulingContext) ClearJobSpecs()
- func (sctx *SchedulingContext) ClearUnfeasibleSchedulingKeys()
- func (sctx *SchedulingContext) EvictGang(jobs []*jobdb.Job) (bool, error)
- func (sctx *SchedulingContext) EvictJob(job *jobdb.Job) (bool, error)
- func (sctx *SchedulingContext) FairnessError() float64
- func (sctx *SchedulingContext) GetQueue(queue string) (fairness.Queue, bool)
- func (sctx *SchedulingContext) ReportString(verbosity int32) string
- func (sctx *SchedulingContext) String() string
- func (sctx *SchedulingContext) SuccessfulJobSchedulingContexts() []*JobSchedulingContext
- func (sctx *SchedulingContext) UpdateFairShares()
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func PrintJobSummary ¶ added in v0.11.0
func PrintJobSummary(ctx *armadacontext.Context, prefix string, jctxs []*JobSchedulingContext)
PrintJobSummary logs a summary of the job scheduling context It will log a high level summary at Info level, and a list of all queues + jobs affected at debug level
Types ¶
type GangInfo ¶ added in v0.4.11
func EmptyGangInfo ¶ added in v0.4.11
func EmptyGangInfo(job interfaces.MinimalJob) GangInfo
EmptyGangInfo returns a GangInfo for a job that is not in a gang.
func GangInfoFromLegacySchedulerJob ¶ added in v0.4.11
func GangInfoFromLegacySchedulerJob(job interfaces.MinimalJob) (GangInfo, error)
type GangSchedulingContext ¶
type GangSchedulingContext struct { Created time.Time Queue string GangInfo JobSchedulingContexts []*JobSchedulingContext TotalResourceRequests schedulerobjects.ResourceList AllJobsEvicted bool RequestsFloatingResources bool }
func NewGangSchedulingContext ¶
func NewGangSchedulingContext(jctxs []*JobSchedulingContext) *GangSchedulingContext
func (*GangSchedulingContext) Cardinality ¶ added in v0.3.92
func (gctx *GangSchedulingContext) Cardinality() int
Cardinality returns the number of jobs in the gang.
func (*GangSchedulingContext) Fit ¶ added in v0.4.2
func (gctx *GangSchedulingContext) Fit() GangSchedulingFit
func (*GangSchedulingContext) JobIds ¶ added in v0.4.27
func (gctx *GangSchedulingContext) JobIds() []string
JobIds returns a sliced composed of the ids of the jobs that make up the gang.
type GangSchedulingFit ¶ added in v0.4.2
type GangSchedulingFit struct { // The number of jobs in the gang that were successfully scheduled. NumScheduled int // The mean PreemptedAtPriority among successfully scheduled pods in the gang. MeanPreemptedAtPriority float64 }
func (GangSchedulingFit) Less ¶ added in v0.4.2
func (f GangSchedulingFit) Less(other GangSchedulingFit) bool
type JobSchedulingContext ¶
type JobSchedulingContext struct { // Time at which this context was created. Created time.Time // Id of the job this pod corresponds to. JobId string // Indicates whether this context is for re-scheduling an evicted job. IsEvicted bool // Job spec. Job *jobdb.Job // Scheduling requirements of this job. // We currently require that each job contains exactly one pod spec. PodRequirements *schedulerobjects.PodRequirements // Resource requirements in an efficient internaltypes.ResourceList ResourceRequirements internaltypes.ResourceList // Node selectors to consider in addition to those included with the PodRequirements. // These are added as part of scheduling to further constrain where nodes are scheduled, // e.g., to ensure evicted jobs are re-scheduled onto the same node. // // If some key appears in both PodRequirements.NodeSelector and AdditionalNodeSelectors, // the value in AdditionalNodeSelectors trumps that of PodRequirements.NodeSelector. AdditionalNodeSelectors map[string]string // Tolerations to consider in addition to those included with the PodRequirements. // These are added as part of scheduling to expand the set of nodes a job can be scheduled on. AdditionalTolerations []v1.Toleration // Reason for why the job could not be scheduled. // Empty if the job was scheduled successfully. UnschedulableReason string // Pod scheduling contexts for the individual pods that make up the job. PodSchedulingContext *PodSchedulingContext // GangInfo holds all the information that is necessary to schedule a gang, // such as the lower and upper bounds on its size. GangInfo // This is the node the pod is assigned to. // This is only set for evicted jobs and is set alongside adding an additionalNodeSelector for the node AssignedNodeId string }
JobSchedulingContext is created by the scheduler and contains information about the decision made by the scheduler for a particular job.
func JobSchedulingContextFromJob ¶ added in v0.4.1
func JobSchedulingContextFromJob(job *jobdb.Job) *JobSchedulingContext
func JobSchedulingContextsFromJobs ¶ added in v0.3.78
func JobSchedulingContextsFromJobs[J *jobdb.Job](jobs []J) []*JobSchedulingContext
func (*JobSchedulingContext) AddNodeSelector ¶ added in v0.4.1
func (jctx *JobSchedulingContext) AddNodeSelector(key, value string)
func (*JobSchedulingContext) Fail ¶ added in v0.4.2
func (jctx *JobSchedulingContext) Fail(unschedulableReason string)
func (*JobSchedulingContext) GetAssignedNodeId ¶ added in v0.8.3
func (jctx *JobSchedulingContext) GetAssignedNodeId() string
func (*JobSchedulingContext) IsSuccessful ¶
func (jctx *JobSchedulingContext) IsSuccessful() bool
func (*JobSchedulingContext) SchedulingKey ¶ added in v0.4.1
func (jctx *JobSchedulingContext) SchedulingKey() (schedulerobjects.SchedulingKey, bool)
SchedulingKey returns the scheduling key of the embedded job. If the jctx contains additional node selectors or tolerations, the key is invalid and the second return value is false.
func (*JobSchedulingContext) SetAssignedNodeId ¶ added in v0.8.3
func (jctx *JobSchedulingContext) SetAssignedNodeId(assignedNodeId string)
func (*JobSchedulingContext) String ¶
func (jctx *JobSchedulingContext) String() string
type PodSchedulingContext ¶
type PodSchedulingContext struct { // Time at which this context was created. Created time.Time // ID of the node that the pod was assigned to, or empty. NodeId string // If set, indicates that the pod was scheduled on a specific node type. WellKnownNodeTypeName string // Priority this pod was most recently attempted to be scheduled at. // If scheduling was successful, resources were marked as allocated to the job at this priority. ScheduledAtPriority int32 // Maximum priority that this pod preempted other pods at. PreemptedAtPriority int32 // Total number of nodes in the cluster when trying to schedule. NumNodes int // Number of nodes excluded by reason. NumExcludedNodesByReason map[string]int }
PodSchedulingContext is returned by SelectAndBindNodeToPod and contains detailed information on the scheduling decision made for this pod.
func (*PodSchedulingContext) IsSuccessful ¶ added in v0.4.2
func (pctx *PodSchedulingContext) IsSuccessful() bool
func (*PodSchedulingContext) String ¶
func (pctx *PodSchedulingContext) String() string
type QueueSchedulingContext ¶
type QueueSchedulingContext struct { // The scheduling context to which this QueueSchedulingContext belongs. SchedulingContext *SchedulingContext // Time at which this context was created. Created time.Time // Queue name. Queue string // Determines the fair share of this queue relative to other queues. Weight float64 // Limits job scheduling rate for this queue. // Use the "Started" time to ensure limiter state remains constant within each scheduling round. Limiter *rate.Limiter // Total resources assigned to the queue across all clusters by priority class priority. // Includes jobs scheduled during this invocation of the scheduler. Allocated schedulerobjects.ResourceList // Total demand from this queue. This is essentially the cumulative resources of all non-terminal jobs at the // start of the scheduling cycle Demand schedulerobjects.ResourceList // Capped Demand for this queue. This differs from Demand in that it takes into account any limits that we have // placed on the queue CappedDemand schedulerobjects.ResourceList FairShare float64 AdjustedFairShare float64 // Total resources assigned to the queue across all clusters by priority class. // Includes jobs scheduled during this invocation of the scheduler. AllocatedByPriorityClass schedulerobjects.QuantityByTAndResourceType[string] // Resources assigned to this queue during this scheduling cycle. ScheduledResourcesByPriorityClass schedulerobjects.QuantityByTAndResourceType[string] // Resources evicted from this queue during this scheduling cycle. EvictedResourcesByPriorityClass schedulerobjects.QuantityByTAndResourceType[string] // Job scheduling contexts associated with successful scheduling attempts. SuccessfulJobSchedulingContexts map[string]*JobSchedulingContext // Job scheduling contexts associated with unsuccessful scheduling attempts. UnsuccessfulJobSchedulingContexts map[string]*JobSchedulingContext // Jobs evicted in this round. EvictedJobsById map[string]bool }
QueueSchedulingContext captures the decisions made by the scheduler during one invocation for a particular queue.
func (*QueueSchedulingContext) ClearJobSpecs ¶
func (qctx *QueueSchedulingContext) ClearJobSpecs()
ClearJobSpecs zeroes out job specs to reduce memory usage.
func (*QueueSchedulingContext) GetAllocation ¶ added in v0.3.90
func (qctx *QueueSchedulingContext) GetAllocation() schedulerobjects.ResourceList
GetAllocation is necessary to implement the fairness.Queue interface.
func (*QueueSchedulingContext) GetWeight ¶ added in v0.3.90
func (qctx *QueueSchedulingContext) GetWeight() float64
GetWeight is necessary to implement the fairness.Queue interface.
func (*QueueSchedulingContext) ReportString ¶ added in v0.3.71
func (qctx *QueueSchedulingContext) ReportString(verbosity int32) string
func (*QueueSchedulingContext) String ¶
func (qctx *QueueSchedulingContext) String() string
type SchedulingContext ¶
type SchedulingContext struct { // Time at which the scheduling cycle started. Started time.Time // Time at which the scheduling cycle finished. Finished time.Time // Pool for which we're currently scheduling jobs. Pool string // Determines how fairness is computed. FairnessCostProvider fairness.FairnessCostProvider // Limits job scheduling rate globally across all queues. // Use the "Started" time to ensure limiter state remains constant within each scheduling round. Limiter *rate.Limiter // Sum of queue weights across all queues. WeightSum float64 // Per-queue scheduling contexts. QueueSchedulingContexts map[string]*QueueSchedulingContext // Total resources across all clusters in this pool available at the start of the scheduling cycle. TotalResources schedulerobjects.ResourceList // Allocated resources across all clusters in this pool Allocated schedulerobjects.ResourceList // Resources assigned across all queues during this scheduling cycle. ScheduledResources schedulerobjects.ResourceList ScheduledResourcesByPriorityClass schedulerobjects.QuantityByTAndResourceType[string] // Resources evicted across all queues during this scheduling cycle. EvictedResources schedulerobjects.ResourceList EvictedResourcesByPriorityClass schedulerobjects.QuantityByTAndResourceType[string] // Total number of successfully scheduled jobs. NumScheduledJobs int // Total number of successfully scheduled gangs. NumScheduledGangs int // Total number of evicted jobs. NumEvictedJobs int // TODO(reports): Count the number of evicted gangs. // Reason for why the scheduling round finished. TerminationReason string // Used to efficiently generate scheduling keys. SchedulingKeyGenerator *schedulerobjects.SchedulingKeyGenerator // Record of job scheduling requirements known to be unfeasible. // Used to immediately reject new jobs with identical reqirements. // Maps to the JobSchedulingContext of a previous job attempted to schedule with the same key. UnfeasibleSchedulingKeys map[schedulerobjects.SchedulingKey]*JobSchedulingContext }
SchedulingContext contains information necessary for scheduling and records what happened in a scheduling round.
func NewSchedulingContext ¶
func NewSchedulingContext( pool string, fairnessCostProvider fairness.FairnessCostProvider, limiter *rate.Limiter, totalResources schedulerobjects.ResourceList, ) *SchedulingContext
func (*SchedulingContext) AddGangSchedulingContext ¶
func (sctx *SchedulingContext) AddGangSchedulingContext(gctx *GangSchedulingContext) (bool, error)
func (*SchedulingContext) AddJobSchedulingContext ¶
func (sctx *SchedulingContext) AddJobSchedulingContext(jctx *JobSchedulingContext) (bool, error)
AddJobSchedulingContext adds a job scheduling context. Automatically updates scheduled resources.
func (*SchedulingContext) AddQueueSchedulingContext ¶ added in v0.3.68
func (sctx *SchedulingContext) AddQueueSchedulingContext( queue string, weight float64, initialAllocatedByPriorityClass schedulerobjects.QuantityByTAndResourceType[string], demand schedulerobjects.ResourceList, cappedDemand schedulerobjects.ResourceList, limiter *rate.Limiter, ) error
func (*SchedulingContext) AllocatedByQueueAndPriority ¶
func (sctx *SchedulingContext) AllocatedByQueueAndPriority() map[string]schedulerobjects.QuantityByTAndResourceType[string]
AllocatedByQueueAndPriority returns map from queue name and priority to resources allocated.
func (*SchedulingContext) ClearJobSpecs ¶
func (sctx *SchedulingContext) ClearJobSpecs()
ClearJobSpecs zeroes out job specs to reduce memory usage.
func (*SchedulingContext) ClearUnfeasibleSchedulingKeys ¶ added in v0.3.68
func (sctx *SchedulingContext) ClearUnfeasibleSchedulingKeys()
func (*SchedulingContext) EvictGang ¶
func (sctx *SchedulingContext) EvictGang(jobs []*jobdb.Job) (bool, error)
func (*SchedulingContext) EvictJob ¶
func (sctx *SchedulingContext) EvictJob(job *jobdb.Job) (bool, error)
func (*SchedulingContext) FairnessError ¶ added in v0.11.0
func (sctx *SchedulingContext) FairnessError() float64
FairnessError returns the cumulative delta between adjusted fair share and actual share for all users who are below their fair share
func (*SchedulingContext) GetQueue ¶ added in v0.3.90
func (sctx *SchedulingContext) GetQueue(queue string) (fairness.Queue, bool)
GetQueue is necessary to implement the fairness.QueueRepository interface.
func (*SchedulingContext) ReportString ¶ added in v0.3.71
func (sctx *SchedulingContext) ReportString(verbosity int32) string
func (*SchedulingContext) String ¶
func (sctx *SchedulingContext) String() string
func (*SchedulingContext) SuccessfulJobSchedulingContexts ¶
func (sctx *SchedulingContext) SuccessfulJobSchedulingContexts() []*JobSchedulingContext
func (*SchedulingContext) UpdateFairShares ¶ added in v0.8.3
func (sctx *SchedulingContext) UpdateFairShares()
UpdateFairShares updates FairShare and AdjustedFairShare for every QueueSchedulingContext associated with the SchedulingContext. This works by calculating a far share as queue_weight/sum_of_all_queue_weights and an AdjustedFairShare by resharing any unused capacity (as determined by a queue's demand)