Static Scheduling in Clouds Thomas A. Henzinger Anmol V. Singh - PowerPoint PPT Presentation

Static Scheduling in Clouds Thomas A. Henzinger Anmol V. Singh Vasu Singh Thomas Wies Damien Zufferey IST Austria June 14, 2011

Motivation (1) Cloud computing gives the illusion of ∞ (virtual) resources. Actually there is a finite amount of (physical) resources. We would like to efficiently share those resources: 1 being able to distinguish high priority (serving customer now ) from low priority (batch) requests; 2 schedule accordingly. Therefore, we should be able to plan ahead computations. Damien Zufferey Static Scheduling in Clouds HotCloud’11 2 / 13

Motivation (2) Dynamic Scheduling: use work queues, priorities, but limited. Without knowledge of jobs, this is the best you can do. We need to ask the user for: what kind of resources his job require; a deadline/priority for his job. In exchange we can give him an expected completion time. We can also offer choice. (time is money.) Damien Zufferey Static Scheduling in Clouds HotCloud’11 3 / 13

Flextic Overview Execution Task finish Cloud Job Parser Plan updates Representation Job Execution Platform Program Job Scheduler User Interface Schedules User chosen schedule User Choice Damien Zufferey Static Scheduling in Clouds HotCloud’11 4 / 13

Giving incentive to plan in advance The scheduler returns not one but many possible schedules with different finish times. Use a pricing model to associate a cost to the schedules. Include the “scheduling difficulty” in the cost, give a discount to schedule with later finish time. Price goes to ∞ as time reaches the minimum makespan price Price converges to 0 as time goes to ∞ time Minimum makespan (critical path) Problem: static scheduling is hard . Only possible if the scheduler can handle the work load. Damien Zufferey Static Scheduling in Clouds HotCloud’11 5 / 13

Jobs Model 1.5 7 10 2 t1 t5 1 t2 t6 4 t9 t3 t7 t4 t8 A Job is a directed acyclic task (DAG) of tasks. Node are marked with worst case duration. Edges are marked with data transfer. duration and data can be parametric in the input. Damien Zufferey Static Scheduling in Clouds HotCloud’11 6 / 13

Parametric Jobs User Job Schema Connections Mappers Reducers Database Job Parser Input Data Size Execution Plan Task Details Object Sizes Damien Zufferey Static Scheduling in Clouds HotCloud’11 7 / 13

Infrastructure Model Router Router Router Router Router Router Datacenter as a tree-like graph: internal nodes are router; leaves are compute nodes (computation speed); edges specifies the bandwidth. Damien Zufferey Static Scheduling in Clouds HotCloud’11 8 / 13

Scheduling Large Jobs using Abstraction [EuroSys 2011] Assumption: job and infrastructure regularity Idea: regularity makes large scale scheduling feasible How: Using abstraction techniques Execution Abstract EP Abstraction Plan Cloud Abstract Cloud Abstraction Representation Scheduler Cloud Abs. Schedule Concretization Representation Damien Zufferey Static Scheduling in Clouds HotCloud’11 9 / 13

Abstraction for jobs: Group independent tasks as per a topological sort. Merge them into an abstract task. 1.5 7 10 2 t1 t5 1 1.8 8 12 2 t2 t6 1 1 4 t9 3 12 20 4 t3 t7 1 1 2.4 10 16 3 t4 t8 Damien Zufferey Static Scheduling in Clouds HotCloud’11 10 / 13

Abstraction for jobs: Group independent tasks as per a topological sort. Merge them into an abstract task. 1.5 7 10 2 t1 t5 1 1.8 8 12 2 t2 t6 1 1 4 t9 3 12 20 4 t3 t7 1 1 2.4 10 16 3 t4 t8 Damien Zufferey Static Scheduling in Clouds HotCloud’11 10 / 13

Abstraction for jobs: Group independent tasks as per a topological sort. Merge them into an abstract task. 3 12 1 20 4 1 4 #4 #4 t9 Damien Zufferey Static Scheduling in Clouds HotCloud’11 10 / 13

Abstraction for infrastructure: Merge nodes to according to network topology: 2 3 / 6 2 2 1 1 1 2 1 / 3 2 2 / 3 1 1 1 1 1 1 1 1 3 3 2 3 2 2 busy busy busy Medium Coarsest Concrete System abstraction abstraction Damien Zufferey Static Scheduling in Clouds HotCloud’11 11 / 13

Experiments: compared to Hadoop Caution: static scheduling alone will not work. Task duration are conservative estimates; Variability of the performance of the compute node. We use static scheduling with backfilling. 200 120 Hadoop Hadoop FISCH FISCH BLIND 100 Job duration (in seconds) 150 Normalized job duration 80 100 60 50 40 0 20 1 2 4 6 8 10 10 20 30 40 50 60 70 80 Number of m1.xlarge instances Number of virtual cores The jobs are MapReduce jobs doing image transformation. Hadoop streaming version 0.19.0 Damien Zufferey Static Scheduling in Clouds HotCloud’11 12 / 13

Questions ? Damien Zufferey Static Scheduling in Clouds HotCloud’11 13 / 13