ChEsS: Cost-Effective Scheduling across multiple heterogeneous mapreduce clusters

Transcription

1 Summarized by: Michael Bowen ChEsS: Cost-Effective Scheduling across multiple heterogeneous mapreduce clusters Nikos Zacheilas Vana Kalogeraki 2016 IEEE International Conference on Autonomic Computing

2 Presentation Summary Key Terminology Problem Statement Contributions Challenges Variables Strategy (Methodology) Impact Estimation Optimization Problem Adaptive Weighted Sum Evaluation

3 Key Terminology Mapreduce - Word Count - the canonical example Photo Credit: wikis.nyu.edu

4 Key Terminology Isolation Data Isolation Privacy and security Failure Isolation Hide failures across clusters Version Isolation Dependency and version management Performance Isolation Prod, Dev, Test - multiple prod different delineations

5 Key Terminology Per-job considerations Performance Monetary Cost Data Locality Scheduling Policy FIFO, Fair, and Capacity - more recently EDF and Least-Laxity

6 Key Terminology Makespan End-to-end execution time of the submitted job Pareto-based analysis Many possible courses of action competing for attention

7 Problem Statement

8 Challenges Jobs-to-clusters possible assignments is an exponential increase as number of jobs and clusters increases Difficult to manually determine these assignments Budget required vs workload makespan

9 Contributions Parameter impact estimates Jobs locality constraints, intra-job scheduling algorithms, etc Pareto-frontier search algorithm improvements Budget vs Makespan tradeoff analysis Evaluation study of industry workloads

10 Variables c Clusters j Jobs VMs c # virtual machines mslots c, rslots c mtasks j, rtasks j # map/reduce tasks used by job j map/reduce slots mslots j,c, rslots j,c schedulerc scheduling algorithm # map/reduce slots reserved by j from c costc size j per hour cost (ec2) input data size of j threadsc Jobs c # threads spawned for execution datahost j where input resides set of jobs assigned to cluster c mtime j,c, rtime j,c, stime j,c makespan c budget c total execution time in seconds of all jobs assigned to c JTime j,c map/reduce/shuffle time estimates required budget execution time of job j

11 Strategy 1. Estimate impact of intra-cluster scheduling policies and locality constraints on makespan and budget 2. Formulate multi-objective optimization problem 3. Solve using Adaptive Weighted Sum (AWS)

12 Impact Estimation Key Assumption - repetitive, aperiodic jobs Execution time - Lower Bound Upper Bound Map/Reduce/Shuffle for lower and upper Final estimate - average of two limits

13 Impact Estimation Key Assumption - repetitive, aperiodic jobs Locality Constraints - Add overhead of time to transfer data to execution time Makespan - Simulator Engine Input - scheduling policy and set of jobs Output - makespan Budget - Budget vs Exec Time -

14 Optimization Problem Pareto-frontier Detect optimal solutions with respect to constraints Result helps user decide amongst solution space Example with two job-to-cluster assignments, P and Q Q dominates P if and only if Budget Q Budget P AND Makespan Q < Makespan P Budget Q < Budget P OR Makespan Q < Makespan P The set of non-dominated assignments is the solution space of interest - known as the Pareto-frontier

15 Adaptive Weighted Sum Pareto-frontier search is very costly - use Adaptive Weighted Sum as an approximation Regular weighted sum - Greedy - assign jobs to clusters that lead to min utilityscore Challenge - Detected solutions non-uniformally distributed Cannot detect solutions in non-convex regions of the solution space Adaptive Weighted Sum - Perform single-objective optimization in unexplored regions of the solution space

16 Evaluation Note - developed for INSIGHT, which provides realtime event detection in Dublin Used industry workloads based on Yahoo s Hadoop clusters Used scientific workloads based on traces from Open- Cloud cluster provider Four possible clusters considered for the possible jobs

17 Evaluation Execution time estimation error

18 Evaluation d j parameter impact

19 Evaluation Scheduling Algorithm Impact

20 Evaluation Locality Constraints Impact

21 Evaluation Comparison With Optimal

22 Critique Assumption of repetitive, aperiodic jobs Understandable constraint - difficult to model otherwise Unsure of how realistic this constraint is Mapreduce is more of a legacy system at this point Rapidly losing market-share to Spark