[
https://issues.apache.org/jira/browse/YUNIKORN-3445?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
weichen lai updated YUNIKORN-3445:
----------------------------------
Summary: SortAllocationsBasedOnAsk prioritizes originator and opted-out
pods for preemption instead of protecting them (was: [Core]
SortAllocationsBasedOnAsk prioritizes originator and opted-out pods for
preemption instead of protecting them)
> SortAllocationsBasedOnAsk prioritizes originator and opted-out pods for
> preemption instead of protecting them
> -------------------------------------------------------------------------------------------------------------
>
> Key: YUNIKORN-3445
> URL: https://issues.apache.org/jira/browse/YUNIKORN-3445
> Project: Apache YuniKorn
> Issue Type: Bug
> Components: core - scheduler
> Affects Versions: 1.9.0
> Reporter: weichen lai
> Assignee: weichen lai
> Priority: Major
>
> h3. Overview
> In {{pkg/scheduler/objects/preemption_utilities.go}},
> {{SortAllocationsBasedOnAsk()}} and {{scoreAllocationBasedOnAsk()}} determine
> victim ordering during preemption.
> Under the current implementation, allocations with {{!allowPreemptSelf}}
> (opted-out) or {{isOriginator}} (driver/manager) receive higher-order score
> bits ({{scoreNoPreempt}} = {{1 << 34}}, {{scoreOriginator}} = {{1 << 33}}):
> {code:go}func scoreAllocationBasedOnAsk(allocation *Allocation, ask
> *resources.Resource) uint64 {
> var score uint64 = 0
> if allocation.IsOriginator() {
> score |= scoreOriginator // 1 << 33
> }
> if !allocation.IsAllowPreemptSelf() {
> score |= scoreNoPreempt // 1 << 34
> }
> score += allocation.GetAllocatedResource().TypeMatching(ask)
> return score
> }
> {code}
> Because {{SortAllocationsBasedOnAsk()}} orders allocations in descending
> order ({{scoreLeft > scoreRight}}), opted-out and originator allocations
> receive higher scores and are placed at the beginning of the candidate slice
> rather than at the end.
> h2.
> h3. Comparison with Design Documentation
> This ordering conflicts with documented design specifications:
> * **Preemption Law #1 (preemption policies are strong suggestions, not
> guarantees):** Opting out of preemption and originator status are intended as
> a last resort when no other suitable allocations can be preempted.
> * **Quota Preemption Design Document:** The specification states that
> non-originators and pods with {{allowPreemption: true}} should receive higher
> scores to prioritize regular allocations for preemption over originators and
> opted-out pods.
> h2.
> h3. Test Verification
> The following test demonstrates the ordering in
> {{pkg/scheduler/objects/preemption_utilities_test.go}}:
> {code:go}func TestSortAllocationsBasedOnAsk_PreemptionOrdering(t *testing.T) {
> node := NewNode(&si.NodeInfo{
> NodeID: "node1",
> SchedulableResource: &si.Resource{
> Resources: map[string]*si.Quantity{"first": {Value:
> 100}},
> },
> })
> res :=
> resources.NewResourceFromMap(map[string]resources.Quantity{"first": 10})
> total :=
> resources.NewResourceFromMap(map[string]resources.Quantity{"first": 100})
> ask :=
> resources.NewResourceFromMap(map[string]resources.Quantity{"first": 10})
> regularPod := createAllocation("regularPod", "app1", node.NodeID, true,
> false, 10, false, res)
> originatorPod := createAllocation("originatorPod", "app1", node.NodeID,
> true, true, 10, false, res)
> optedOutPod := createAllocation("optedOutPod", "app1", node.NodeID,
> false, false, 10, false, res)
> allocations := []*Allocation{regularPod, originatorPod, optedOutPod}
> SortAllocationsBasedOnAsk(allocations, total, ask)
> // Current ordering:
> // allocations[0]: optedOutPod
> // allocations[1]: originatorPod
> // allocations[2]: regularPod
> //
> // Expected ordering per design: regularPod should be placed at index 0
> assert.Equal(t, allocations[0].GetAllocationKey(), "regularPod",
> "Expected regularPod at index 0, but got: %s",
> allocations[0].GetAllocationKey())
> }
> {code}
> *Output:*
> {code}=== RUN TestSortAllocationsBasedOnAsk_PreemptionOrdering
> preemption_utilities_test.go:320: Expected regularPod at index 0, but
> got: optedOutPod
> --- FAIL: TestSortAllocationsBasedOnAsk_PreemptionOrdering (0.00s)
> {code}
> h2.
> h3. Impact
> During Quota Preemption, reducing a queue's quota results in originator and
> opted-out allocations being preempted before standard worker allocations.
> h2.
> h3. Proposed Solution
> In {{pkg/scheduler/objects/preemption_utilities.go}}, assign the higher-order
> score bits to unprotected allocations ({{allowPreemptSelf == true}} and
> {{!isOriginator}}):
> {code:go}const (
> scoreAllowPreempt uint64 = 1 << 34
> scoreNonOriginator uint64 = 1 << 33
> )
> func scoreAllocationBasedOnAsk(allocation *Allocation, ask
> *resources.Resource) uint64 {
> var score uint64 = 0
> if !allocation.IsOriginator() {
> score |= scoreNonOriginator
> }
> if allocation.IsAllowPreemptSelf() {
> score |= scoreAllowPreempt
> }
> score += allocation.GetAllocatedResource().TypeMatching(ask)
> return score
> }
> {code}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
