RATATOSKRATATOSK
로그인

Volcano

v1.14.0Orchestration & Management
2026년 1월 31일

OTHER 61

Volcano v1.14.0은 기능 확장과 유지 보수를 함께 담은 릴리스입니다. 여러 기능의 추가와 정확성 및 안정성 문제 수정이 포함됐으며, 보안 권고나 명시된 취약점은 언급되지 않았습니다.

그 외 기록된 변경 61건 전체additions 28 · fixes 25 · constraints 7 · value changes 1

additions (28)

  • The Sharding Controller introduces a scalable multi-scheduler architecture that dynamically computes candidate node pools for each scheduler.
  • Currently supports scheduling shard by CPU utilization, with an extensible design to support more policies in the future.
  • --shard-sync-period=60s --enable-node-event-trigger=true
  • we introduce a dedicated Agent Scheduler.
  • Support highestTierName for specifying maximum network tier constraints by name.
  • Enable partitioning of Volcano Jobs for better resource management and fault isolation.
  • CPU Throttling (CPU Suppression)
  • This feature introduces the ColocationConfiguration CRD, which allows users to define memory QoS policies for specific workloads.
  • introduces integrated support for Ascend vNPU (virtual NPU) scheduling
  • deviceshare.AscendMindClusterVNPUEnable: true
  • deviceshare.AscendHAMiVNPUEnable: true deviceshare.SchedulePolicy: binpack
  • Node Scheduling Shard CRD. Manages dynamic candidate node pools for specific schedulers.
  • Multi-cluster training job abstraction for the Volcano Global enhancements.
  • Data-aware scheduling for federated environments.
  • Defines fine-grained network topology constraints and gang scheduling policies at the SubGroup level.
  • New configuration for queue dequeue strategies.
  • Optional enhanced admission mechanisms
  • Added GHCR as an alternative to Docker Hub
  • Support for updating pods without recreation
  • Enables role-based elasticity and Gang Scheduling for distributed inference workloads
  • **NodeShard CRD**: Manages dynamic candidate node pools for specific schedulers.
  • **Fast-Path Scheduling**: Independent scheduler optimized for latency-sensitive workloads
  • Multiple workers process pods concurrently from the scheduling queue, increasing throughput
  • Conflict-Aware Binder resolves scheduling conflicts before executing real binding
  • Enhanced queue mechanism with urgent retry support
  • **HyperNode-Level Binpacking**: Optimization for resource utilization across network topology boundaries.
  • **MindCluster Mode** - Integrated from the official Ascend cluster scheduling add-on
  • **HAMi Mode** - Developed by the HAMi community

fixes (25)

  • Fixed reclaimableFn to correctly handle scalar resources and ensure victims are selected based on all relevant resource dimensions
  • Addressed edge cases in resource calculation to prevent scheduling loops and incorrect evictions.
  • Fixed timeout due to un-released PreBind lock
  • Fixed panic when child queue capability check fails
  • Fixed job pending state caused by minAvailable/minPartitions mismatch
  • Fixed early transition to Aborted when terminating pods are removed
  • Fixed memory leak caused by leftover metrics data due to race conditions
  • Fixed cluster-wide scheduling failure due to missing validation for child queue guarantee sum
  • Fixed unexpected behavior when pod is evicted
  • Fixed scheduling failures caused by task member changes in PodGroup
  • Fixed memory leak in volcano-scheduler deletedJobs queue
  • Fixed VGPU request when multiple cards have different total video memory
  • Fixed gpu resource error at allocate state
  • Fixed resource comparison preventing reclaim
  • Fixed queue share calculation for best-effort queues
  • Fixed systemd driver with containerd runtime oversubscribed cgroups
  • Fixed panic when numatopology is added and updated in short time before snapshot
  • Fixed inaccurate message when podgroup is ready
  • Fixed nil pod when AlreadyExists error on create
  • Fixed vNPU not working in multi-type NPU clusters
  • Fixed memory allocation restrictions in Hami vNPU scene
  • Fixed multiple bugs in MindCluster vNPU mode
  • Fixed annotation handling for replicaset
  • Fixed to add node futureIdle resource first
  • Fixed cache communication issues

constraints (7)

  • Support fine-grained network topology constraints at the SubGroup/Partition level.
  • Improved gang scheduling to support both Job-level and SubGroup-level consistency.
  • Extended CPU Burst support to generic operating systems (Ubuntu, CentOS, etc.)
  • Volcano Agent now fully supports Cgroup V2 environments
  • Added CEL validation markers to batch, scheduling, topology, and flow APIs.
  • percentage-based coefficients (range 0-100) for highRatio, lowRatio, and minRatio
  • support for generic operating systems

value changes (1)

  • Support Kubernetes 1.34
Volcano 스택에 추가

조치가 필요한 변경이 나오면 주간 메일로 받아볼 수 있습니다. 이번 릴리스처럼 조치가 없는 주에는 메일을 보내지 않습니다.

스택에 추가 →
공식 릴리스 노트 ↗Volcano 릴리스 전체 →