Experiences With Scheduling and Mapping Games for Adaptive Distributed Systems

1. Experiences With Scheduling and MappingGames for Adaptive Distributed Systems Bin Lin, Peter DindaDepartment of EECSNorthwestern University {binlin365, pdinda}@gmail.com empathicsystems.org Game Interface • Is it possible to map a scheduling and mapping problem in distributed and parallel systems to a game that a naïve user can play, with the side effect having of good game play correspond to a good solution to the problem? YES! • Can naive users play such a game well? YES! • Our technique: Human-driven Search • A game for naïve users in which game-play corresponds to solving a well posed, but difficult to solve optimization problem in adaptive virtualized computing • Problem:Maximize the performance of a BSP program running in a collection of VMs • through VM migration and selection of periodic real-time schedules • A physical host – a resource box; A VM – a ball • VM efficiency - happiness of a ball: (% of available compute time being used) • Objective function f(x) - score: global cumulative happiness of all of the balls, • assuming all the balls are working together to make progress towards a global goal (parallel efficiency). • Goal of the game: to achieve the highest possible score. • How to play: play by dragging balls within boxes (VM schedule change), or • between boxes (VM to host mapping). As the user drags a ball, the game highlights and enforces the scheduling and capacity constraints on where the user may place it. • Final screen: global cumulative happiness (f(x)) and its time average The user is trying to migrate the ball (VM) in the left-most resource box (host) to the second resource box (host) to the left. The position in the box (host) corresponds to a periodic real-time schedule for the ball (VM) User Study • 21 users with various backgrounds • 2 warm-up tasks and 9 formal tasks • Conclusions • Considerable variation in user performance • as expected in any game. • In almost all scalesand types of tasks considered, • there are users who perform near-optimally (compared with optimal solutions either by construction or by simulation-based search) • Most users are able to find optimal mappings. • In the worst case task, more than 65% • of users were able to find the optimal VM mapping. • As task difficulty and problem size grow, • the average time to find the optimal mapping grows. However, users were able to find an optimal mapping in 2–3 minutes. Percentage of users who find the optimal mapping; 95% confidence interval. Duration to the optimal mapping Details can be found in Lin’s Dissertation (NWU-EECS-07-04), available from our web site. The Empathic Systems Project (empathicsystems.org) is funded by NSF CNS-0720691