Scheduling Componentised Applications on a Computation Grid

1. Scheduling Componentised Applications on a Computation Grid Laurie Young Transfer Presentation London e-Science Centre Department of Computing, Imperial College London

2. Component Applications Mesh Generator DRACS Design Generator DRACS Mesh Generator Factory Analyser Mesh Generator DRACS • Each job is composed of multiple components. • Each runs on a (potentially) different resource • Each component is connected to at least one other component. • Data is passed along these connections Laurie Young – PhD Transfer

3. What is a Grid • Remote Cluster • Managed externally • Connected via Internet • Local Cluster • Managed locally • Connected directly Laurie Young – PhD Transfer

4. What is Scheduling Mesh Generator DRACS Design Generator DRACS Mesh Generator Factory Analyser Mesh Generator DRACS • The act of placing components onto resources • Each certain requirements might have to be met • Need to find the “best” mapping of components onto resources Laurie Young – PhD Transfer

5. Outline • Grid Architecture • ICENI • Scheduling Architecture • Scheduling Algorithms • Experiments and Results • Where does this fail? • How to proceed… Laurie Young – PhD Transfer

6. Grid Architecture Application Toolkits (Problem Description and User Interface) Nimrod/G Condor-G Fabric Services (Resource Orchestration Software) OGSA Legion Globus Fabric (Resource Management Software) LSF SGE PBS Condor Resources (Physical Hardware Resources) Windows Linux HPC Solaris Laurie Young – PhD Transfer

7. ICENI • ICe-Science Networked Infrastructure • Developed by LeSC Grid Middleware Group • Collect and provide relevant Grid meta-data • Use to define and develop higher-level services • Interaction with other frameworks: OGSA, Jxta etc. The Iceni, under Queen Boudicca, united the tribes of South-East England in a revolt against the occupying Roman forces in AD60. Laurie Young – PhD Transfer

8. ICENI Architecture Fabric (Resource Management Software) LSF SGE PBS Condor Resources (Physical Hardware Resources) Windows Linux HPC Solaris Application Toolkits (Problem Description and User Interface) Nimrod/G Condor-G Fabric Services (Resource Orchestration Software) OGSA Legion Globus Laurie Young – PhD Transfer

9. ICENI Scheduling Architecture ICENI Scheduling Services Launching Framework Pluggable Launchers (SGE, Globus, Condor, ICENI) • Scheduling Framework • Pluggable Schedulers (Simulated Annealing, Game Theory • Random, Best of n Random) • Performance Framework • Pluggable Performance Repositories • (Perf. Models, • Statistical Analysis) Laurie Young – PhD Transfer

10. Schedule Evaluation • Need a way of numerically evaluating a schedule. • Define a benefit function • Fast solutions are good – Time Optimisation • Cheap solutions are good – Cost Optimisation • Fast and cheap solutions are good – but how to model this? ? or Laurie Young – PhD Transfer

11. Random / Best of n Random • Random Scheduler • Randomly selects a schedule • Checks schedule can be executed • Produces schedules very quickly • Best of n Random • Produces multiple random schedules • Returns the best one • Still very fast • Better results than the random schedules Laurie Young – PhD Transfer

12. Simulated Annealing • Monte Carlo method • Generate schedule at random • Modify current schedule • Accept new schedule if better • If worse, accept with probability proportional to “temperature” and inversely proportional to benefit change • Repeat, while reducing “temperature” • Stop when no modifications to schedule accepted Laurie Young – PhD Transfer

13. Game Theory • Each component is a “Player” • Each player has to choose best strategy (Grid resource) • Each strategy has a benefit, depending on the strategy chosen by all other players. • Players identify, then remove strategies guaranteed to never be optimal – “strictly dominated strategies” • Produces the “Nash Equilibrium” Laurie Young – PhD Transfer

14. Experiments • Schedulers • Random / Best of n RandomProduces usable schedules fast. • Game Theory Considers the scheduling problem as an economic problem. • Simulated AnnealingAlgorithm for solving optimisation Problems • Scheduling Policy • Time OptimisationBest benefit from a schedule with the shortest execution time. Results show scheduling time + execution time. • Cost OptimisationBest benefit from a schedule where the cost of using resources is low. • Simulated Scheduling • Framework • Consistent InterfaceUses the same interface as the ICENI scheduling framework allowing the same schedule code to be used. • RepeatabilityAs the underlying description files never change the same experiment can be run many times. • Application Description • 21 Total Applications in DAG configurations • 3 of particular interest • Serial: • 4 components total • 3 components long • Medium • 8 components total • 3 components long • Parallel • 12 components total • 3 components long • Grid Description • 4 Clusters of resources • Saturn16 Sparc III 750 MHz Processors5Gbit Interconnects • Rhea8 Sparc III 900Mhz Processors5Gbit Interconnects • Viking T16 node, 2GHz Pentium 41Gbit Interconnects • Viking C16 node, 2GHz Pentium 4100Mbit Interconnects Laurie Young – PhD Transfer

15. Results (Cost Optimisation) Laurie Young – PhD Transfer

16. Results (Time Optimisation) Laurie Young – PhD Transfer

17. What to do next… Laurie Young – PhD Transfer

18. Identify Problems • Unknown nature of Grid systems • It is very difficult to know the current state of the Grid we are trying to schedule on • Complex Benefit Functions • Non terminating applications… • Synchronous execution • Greedy Scheduling • What stops first application submitted from being given all available resources Laurie Young – PhD Transfer

19. Difficult Grid Properties • Short lifetime of data • Data such as current load of a remote machine is out of date before it can be communicated to the scheduler. Sometimes we might not even know any value. • Stochastic nature of Grids • Fundamentally variables such as job future job arrival rate, and uptime of resources are stochastic • We need to combine last known state with historical trends to predict current values Laurie Young – PhD Transfer

20. Utility • Allows a preferential ordering of schedules • Is in some way dependent on the amount of various “goods” allocated to a component We want 5 resources, of moderate power or greater with high bandwidth between them. We need them for 3 hours each, and they must be allocated simultaneously. If the bandwidth from our data to the resources is slow, we need more time to copy the data We want a large number of processors for a long period of time, arranged such that our data can be sent to any processor quickly, and so that the processors will remain available for the duration of the allocation Laurie Young – PhD Transfer

21. Utility Functions • A utility function is a mathematical expression of the utility • Similar to a benefit function, but contains more abstract terms. • Issue: We are given computer resources as an allocation in time. Not only does the amount of compute power need to be taken into account, when its available is also important! Laurie Young – PhD Transfer

22. Group Level Policy – Collective Utility • We can examine the behaviour of organisations. • Define their utility to be a function of the utilities of individuals contained within the organisation. • Different functions leads to different social policy • Utilitarian – Maximise sum all utilities. • Rawlsian – Maximise the minimum utility • Egalitarian – Minimise the difference in utilities • This is a social choice, not an optimising choice Laurie Young – PhD Transfer

23. Major Issues • Identification of qualitative issues • Grid Data: • Deterministic • Stochastic • Users: • Single user • Multi user • Mapping to quantitative functions • Application of existing stochastic solutions • Implement sample cases Laurie Young – PhD Transfer

24. Thesis Outline • Introduction • Simple scheduling • Utility function scheduling • Collective utility and group policies • Scheduling algorithms, taking into account stochastic variables • Experimental Results • Conclusion Laurie Young – PhD Transfer

25. Timeline Laurie Young – PhD Transfer

26. Acknowledgements • Director:Professor John Darlington • Technical Director: Dr Steven Newhouse • Research Staff: • Anthony Mayer, Nathalie Furmento, Stephen McGough • William Lee, Marko Krznaric, Murtaza Gulamali • Asif Saleem, Laurie Young, Gary Kong, Jeffrey Hau • Angela 0’Brien, Jeremy Cohen, Ali Afzal • Support Staff: • System: • Keith Sephton, David McBride • Operations: • Susan Brooks, Oliver Jevons • Contact: • E-mail: lesc@ic.ac.uk • Web: http://www.lesc.imperial.ac.uk Laurie Young – PhD Transfer

27. The End Thank you for listening Questions Laurie Young – PhD Transfer