GIFT: A Coupon Based Throttle-and-Reward Mechanism for Fair and Efficient I/O Bandwidth
summary
GIFT implements a coupon-based throttle-and-reward system that throttles some jobs and later redeems coupons to restore fairness, using an LP optimizer that enforces synchronous I/O progress and per-OST constraints. A FUSE prototype on 64 OSTs 192 OSCs 250 apps and larger simulations show about +15% effective bandwidth, ~10% I/O time reduction and 2.5% run-time gain while issuing <0.08% extra node-hours in coupons, so the mechanism measurably cuts wasted bandwidth and preserves long-term fairness.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:24The Key Idea Behind
- 10:10Per-OST Fair Share
- 13:17Three Key Ingredients
- 17:41Careful Coupon Redemption GIFT redeems coupons only when it does not require throttling other applications.
- 21:50Optimal I/O Bandwidth Allocation How much to throttle and whom to
- 22:52More GIFT Design and Implementation Details Mathematical formulation of throttle
- 25:04GIFT improves system I/O bandwidth, mean app IO time and runtime
- 29:24GIFT's system regret budget needed to award outstanding hours is low
related talks
21mSvalinn: Overload Control in Large-Scale Servers with Multiple Resource Bottlenecks
25mAvoiding Overload in Distributed Systems
17mKarma: Resource Allocation for Dynamic Demands
16mSöze: One Network Telemetry Is All You Need for Per-flow Weighted Bandwidth Allocation...
25mApproximating Fair Queueing on Reconfigurable Switches
17mEnabling Users to Control their Internet
16mTowards Optimal Rack-scale μs-level CPU Scheduling through In-Network Workload...
40mEverything Will Flow