Hedger Lai created YUNIKORN-3460:
------------------------------------
Summary: Minimize victim pods in k8shim preemption by introducing
reprieve verification
Key: YUNIKORN-3460
URL: https://issues.apache.org/jira/browse/YUNIKORN-3460
Project: Apache YuniKorn
Issue Type: Improvement
Components: shim - kubernetes
Reporter: Hedger Lai
Assignee: Hedger Lai
h3. Problem
In k8shim, {{PreemptionFilter()}} removes candidate victim pods sequentially
starting from {{startIndex}}. Once filter plugins pass at index {{i}}, all
victims from index {{0}} to {{i}} are returned to Core and preempted. This
leads to over-preemption because earlier candidate pods that were not the
actual cause of predicate failures (e.g., {{PodAntiAffinity}}) are
unnecessarily evicted alongside the actual violator.
h3. Root Cause
Kubernetes {{kube-scheduler}} implements a two-phase preemption algorithm in
{{selectVictimsOnNode}}: a forward search phase followed by a reprieve phase
({{reprievePod}}) that attempts to add back victims one by one to ensure only
the minimal required set is evicted.
In YuniKorn, {{predicateManagerImpl.PreemptionFilter()}} only implements the
forward search phase:
{code:go}
for i := startIndex; i < len(victims); i++ {
p.removePodFromNodeNoFail(preemptingNode, victims[i])
status, _ := p.runFilterPlugins(ctx, *p.allocationFilters, cycleState, pod,
preemptingNode)
if status.IsSuccess() {
return i
}
}
{code}
Furthermore, {{si.PreemptionPredicatesResponse}} only returns a scalar {{Index
int32}}, forcing Core to assume all victims in {{[0, Index]}} must be preempted.
h3. Fix
1. Extend {{si.PreemptionPredicatesResponse}} to optionally return {{repeated
string victimAllocationKeys}} (retaining {{Index}} for backward compatibility).
2. In k8shim {{PreemptionFilter()}}, after predicates pass at index {{i}},
introduce a reprieve pass to add back pods one by one. Pods whose restoration
still allows predicates to pass will be retained.
3. Return the minimal set of victim allocation keys to Core.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
