DRF

Dominant Resource Fairness (DRF) allocates multi-resource capacity by equalizing users’ dominant shares, extending scalar fair sharing to workloads that consume different mixes of CPU, memory, storage, or accelerators.

核心思想

A user’s dominant share is its largest normalized demand across resource dimensions. DRF seeks allocations that prevent one workload’s dominant bottleneck from monopolizing a shared cluster, but the result depends on declared demand vectors, resource granularity, placement constraints, and whether resources are divisible or time-varying.

为什么重要

Storage and systems schedulers often face multi-dimensional contention. DRF is a useful fairness reference, but it does not by itself optimize locality, tail latency, deadlines, fragmentation, or application-level value.

关键观察 / 隐含假设

• 观察：fairness objectives can conflict with device or data placement. HARE-FAST26 studies a storage-system scheduling context where resource behavior matters.
• 观察：runtime constraints can make static demand vectors incomplete. Spirit-SOSP25 uses a system-management context with richer operational boundaries.
• 假设：resource demands are observable and stable enough for a fair-share policy to act on them; bursty or phase-changing workloads weaken this assumption.

设计空间与取舍

• Fair share vs efficiency/locality：equal dominant shares can leave capacity unusable when placement or affinity is constrained.
• Static vs adaptive demand：adaptive models respond to workload change but add measurement and policy complexity.
• Cluster resource fairness vs SLOs：a fair allocation may still violate latency/deadline objectives.

引用本概念的论文

• HARE-FAST26 — multi-resource/storage scheduling context.
• Spirit-SOSP25 — operational resource-management context.