Problem Description Revenue QoS Health Power Migration
ILPJob allocation in data-grid• Power consumption based on used CPUs• CPUs in each host• Min CPUs required by each job• Max CPUs required by each job
ILPObjective Function Benefit ofMax: Execution QoS Penalty Power Consumption Migration Cost
ILPS.T: Processor switched on/off in order: keep consistency Relaxation: job scheduled or not scheduled Available CPUs in each host not exceedOutput: Max. Benefit Placement of each job in the infrastracture CPU assignment for each job CPUs used in each host
Data GenerationGenerate an array of numHosts components: cpus: CPUs in each host, each with 1, 2, 4 or 8 CPUs (random).Generate two arrays of numJobs components: consMin: minimum CPU required, between 1 and 10 (random). consMax: maximum CPU required, randomly between consMin[j] + 1 to 2 extra CPUs (random).
CPU : Intel i7 @ 2.8 GHzOS: Windows 7RAM: 8 GBCPLEX: IBM ILOG CPLEX Optimization Studio 12.4Heuristic: Java in JRE 1.6.0_24-b07
Problem Size vs Methodology vs Benefit 700 CPLEX 600 500 Heuristic Non- Random Initial 400 Selection (NR)Benefit Heuristic Random 300 Initial Selection(R) - 10000 Iter 200 Heuristic Random Initial Selection(R) - 100 100000 Iter 0 Problem Size
ConclusionsDatacenter job scheduling and management can be optimized using ILPs.Complex ILP restrictions can be translated into easy heuristic code.CPLEX does not scale well.Heuristics can cope with higher problem sizes.
ConclusionsLower alpha values achieve better results. Alpha of 0 is the best when using random node selection.Random node selection obtains the best results.More iterations achieve better benefits.
ReferenceJ. L. Berral García, R. Gavaldà Mestre, J. TorresViñals, and others, “An integer linearprogramming representation for data-centerpower-aware management,” 2011.http://upcommons.upc.edu/handle/2117/11061
ILP model and HeuristicAuthors: Josep Subirats Arinto Murdopo Ioanna Tsalouchidou