Beyond Dominant Resource Fairness David Parkes (Harvard) Ariel Procaccia (CMU) Nisarg Shah (CMU) Motivation • Allocation of multiple resources (e.g., CPU, RAM, bandwidth) • Users have heterogeneous demands • Today:
Download ReportTranscript Beyond Dominant Resource Fairness David Parkes (Harvard) Ariel Procaccia (CMU) Nisarg Shah (CMU) Motivation • Allocation of multiple resources (e.g., CPU, RAM, bandwidth) • Users have heterogeneous demands • Today:
Beyond Dominant Resource Fairness David Parkes (Harvard) Ariel Procaccia (CMU) Nisarg Shah (CMU) Motivation • Allocation of multiple resources (e.g., CPU, RAM, bandwidth) • Users have heterogeneous demands • Today: fixed bundles (slots) • Allocate slots using single resource abstraction 2 The DRF mechanism • Assume proportional demands (a.k.a. Leontief preferences) • Example: o o o User wishes to execute multiple that requires 2 CPU and 1 RAM Indifferent between 5 CPU and 2 RAM, and 2 GB Happier with 4.2+2.1 • Dominant resource fairness [Ghodsi et al. 2011]: equalize largest shares 3 DRF animated User 1 alloc. User 2 alloc. Total alloc. 4 Properties of DRF • Pareto optimality • Envy freeness: users do not want to swap allocations • Sharing incentives (a.k.a. fair share, proportionality, IR): users receive at least as much value as an equal split • Strategyproofness: reporting true demands is a dominant strategy • Exciting application of fair division theory! 5 Indivisible tasks • Demands specified as fraction of resource r that user i needs to run one instance of its task • User’s utility strictly increases with number of complete instances of task 6 PO+SI+SP are incompatible User 1 demand User 2 demand Allocation User 1 demand User 2 demand Allocation 7 Envy freeness • PO and EF are trivially incompatible • Need to relax the notion of envy freeness [Budish 2011, Lipton et al. 2004, Moulin and Stong 2002] • Envy freeness up to one bundle (EF1) = i does not prefer j’s after removing one copy of i’s task • Theorem: PO+EF1+SP impossible 8 Sequential Minmax • • • • SI+EF1+SP trivial SI+PO+SP, EF1+PO+SP impossible Can we achieve PO+SI+EF1? The SEQUENTIAL MINMAX mechanism: allocate at each step to minimize maximum allocated share after allocation • Theorem: Mechanism is PO+SI+EF1 9 Sequential Minmax illustrated User 1 demand User 1 alloc. User 2 demand User 2 alloc. Total alloc. 10 Discussion • Additional results in paper o o An extension of DRF to settings zero demands and endowments, satisfies group strategyproofness Lower bounds on social welfare maximization • Current work: dynamic fairness 11