1. New Development in the AppLeS Project or User-Level Middleware for the Grid Francine Berman University of California, San Diego

2. The Evolving Grid Applications Resources In the beginning, there were applications and resources, and it took ninja programmers and many months to implement the applications on the Grid …

3. The Evolving Grid Grid Middleware Applications Resources Applications Resources And behold, there were services, and programmers saw that it was good (even though their performance was still often less than desirable) …

4. The Evolving Grid Grid Middleware Applications Resources User-Level Middleware Grid Middleware Applications Resources Applications Resources … and it came to pass that user-level middleware was promised to promote the performance of Grid applications, and the users rejoiced …

5. The Middleware Promise Grid Middleware –Provides infrastructure/services to enable usability of the Grid –Promotes portability and retargetability User-level Middleware –Hides the complexity of the Grid for the end-user –Adapts to dynamic resource performance variations –Promotes application performance Grid Middleware Applications Resources User-Level Middleware

6. How Do Applications Achieve Performance Now? AppLeS = Application-Level Scheduler –Joint project with R. Wolski –AppLeS + application = self- scheduling Grid application –AppLeS-enabled applications adapt to dynamic performance variations in Grid Resources Grid Middleware Resources AppLeS- enabled applications

7. AppLeS Architecture Grid Middleware Resources AppLeS- enabled applications Schedule Deployment Resource Discovery Resource Selection Schedule Planning and Performance Modeling Decision Model accessible resources feasible resource sets evaluated schedules “best” schedule

8. From AppLeS-enabled applications to User-Level Middleware Grid Middleware Applications Resources User-Level Middleware Grid Middleware Resources AppLeS- enabled applications AppLeS agent integrated within application

9. AppLeS User-Level Middleware Focus is development of templates which –target structurally similar classes of applications –can be instantiated in a user-friendly timeframe –provide good application performance Application Module Scheduling Module Deployment Module Grid Middleware and Resources AppLeS Template Architecture

10. APST – AppLeS Parameter Sweep Template Parameter Sweeps = class of applications which are structured as multiple instances of an “experiment” with distinct parameter sets Joint work with Henri Casanova First AppLeS Middleware package to be distributed to users Parameter Sweeps are common application structure used in various fields of science and engineering –Most notably: Simulations (Monte Carlo, etc.) Large number of tasks, no task precedences in the general case  easy scheduling ? –I/O constraints –Need for meaningful partial results –multiple stages of post-processing

11. APST Scheduling Issues Large shared files, if any, must be stored strategically Post-processing must minimize file transfers Adaptive scheduling necessary to account for changing environment

12. Contingency Scheduling: Allocation developed by dynamically generating a Gantt chart for scheduling unassigned tasks between scheduling events Basic skeleton 1.Compute the next scheduling event 2.Create a Gantt Chart G 3.For each computation and file transfer currently underway, compute an estimate of its completion time and fill in the corresponding slots in G 4.Select a subset T of the tasks that have not started execution 5.Until each host has been assigned enough work, heuristically assign tasks to hosts, filling in slots in G 6.Implement schedule Scheduling Approach 1 2 1 2 1 2 Network links Hosts (Cluster 1) Hosts (Cluster 2) Time Resources Computation G Scheduling event Scheduling event Computation

13. Scheduling Heuristics Self-scheduling Algorithms workqueue workqueue w/ work stealing workqueue w/ work duplication... Gantt chart heuristics: MinMin, MaxMin Sufferage, XSufferage... Scheduling Algorithms for PS Applications Easy to implement and quick No need for performance predictions Insensitive to data placement More difficult to implement Needs performance predictions Sensitive to data placement Simulation results (HCW ’00 paper) show that: heuristics are worth it Xsufferage is good heuristic even when predictions are bad complex environments require better planning (Gantt chart)

14. NetSolve Globus Legion NWS Ninf IBP Condor APST Architecture transport APIexecution API metadata API scheduler API Grid Resources and Middleware APST Daemon GASSIBP NFS GRAMNetSolve Condor, Ninf, Legion,.. NWS Workqueue Gantt chart heuristic algorithms Workqueue++ MinMinMaxMinSufferageXSufferage APST Client Controller interacts Command-line client Metadata Bookkeeper Actuator Scheduler triggers transferexecutequery store actuate report retrieve

15. APST APST being used for –INS2D (NASA Fluid Dynamics application) –MCell (Salk, Molecular modeling for Biology) –Tphot (SDSC, Proton Transport application) –NeuralObjects (NSI, Neural network simulations) –CS simulation Applications for our own research (Model validation, long-range forecasting validation) Actuator’s APIs are interchangeable and mixable –(NetSolve+IBP) + (GRAM+GASS) + (GRAM+NFS) Scheduler API allows for dynamic adaptation No Grid software is required –However lack of it (NWS, GASS, IBP) may lead to poorer performance More details in SC’00 paper

16. APST Validation Experiments University of Tennessee, Knoxville NetSolve + IBP University of California, San Diego GRAM + GASS Tokyo Institute of Technology NetSolve + NFS NetSolve + IBP APST Daemon APST Client

17. APST Test Application – MCell MCell = General simulator for cellular microphysiology Uses Monte Carlo diffusion and chemical reaction algorithm in 3D to simulate complex biochemical interactions of molecules Focus of new multi- disciplinary ITR project –Will focus on large-scale execution-time computational steering, data analysis and visualization

18. Experimental Results Experimental Setting: Mcell simulation with 1,200 tasks: composed of 6 Monte-Carlo simulations input files: 1, 1, 20, 20, 100, and 100 MB 4 scenarios: Initially (a) all input files are only in Japan (b) 100MB files replicated in California (c) in addition, one 100MB file replicated in Tennessee (d) all input files replicated everywhere workqueue Gantt-chart algs

19. New Directions: “Mega-programming” Grid programs –Can reasonably obtain some information about environment (NWS predictions, MDS, HBM, …) –Can assume that login, authentication, monitoring, etc. available on target execution machines –Can assume that programs run to completion on execution platform Mega-programs –Cannot assume any information about target environment –Must be structured to treat target device as unfriendly host (cannot assume ambient services) –Must be structured for “throwaway” end devices –Must be structured to run continuously

20. Success with Mega- programming Seti@home –Over 2 million users –Sustains teraflop computing Can we run non-embarrassingly parallel codes successfully at this scale? –Computational Biology, Genomics … –Genome@home

21. Genome@home Joint work with Derrick Kondo Application template for peer- to-peer platforms First algorithm (Needleman- Wunsch Global Alignment) uses dynamic programming Plan is to use template with additional genomics applications Being developed for “web” rather than Grid environment GTAAG A 00110 T 01011 A 002 21 C 00122 C 00122 G 10123 Optimal alignments determined by traceback

22. Mega-programs Provide the algorithmic counterpart for very large scale platforms –peer-to-peer platforms, Entropia, etc. –Condor flocks –Large “free agent” environments –Globus –New platforms: networks of low-level devices, etc. Different computing paradigm than MPP, Grid Globus Legion free agents … Entropia Condor Algorithm2 DNA Alignment Algorithm1 Genome@home

23. Coming soon to a computer near you: –Release of APST v0.1 by SC’00 –Release of AMWAT (AppLeS Master/ Worker Application Template) v0.1 by Jan ‘01 –First prototype of genome@home: 2001 genome@home –AppLeS software and papers: http://apples.ucsd.edu Thanks! –NSF, NPACI, NASA Grid Computing Lab: –Fran Berman (berman@cs.ucsd.edu) –Henri Casanova –Walfredo Cirne –Holly Dail –Marcio Faerman –Jim Hayes –Derrick Kondo –Graziano Obertelli –Gary Shao –Otto Sievert –Shava Smallen –Alan Su –Renata Teixeira –Nadya Williams –Eric Wing –Qiao Xin

24. Scheduling Results [1] Heuristics for Scheduling Parameter Sweep Applications in Grid Environments H. Casanova, A. Legrand, D. Dzagorodnov, F. Berman (HCW’00) Self-scheduling Algorithms workqueue workqueue w/ work stealing workqueue w/ work duplication... Algorithms using Gantt charts: (using heuristics) MinMin, MaxMin Sufferage, XSufferage... Scheduling Algorithms for PS Applications ? Easy to implement and quick No need for performance predictions Extremely adaptive No planning (resource selection, I/O, …) More difficult to implement Slower to run Needs performance predictions Tunable adaptivity Heuristics for better planning Simulation results in [1] show that: heuristics are worth it Xsufferage is good heuristic even when predictions are bad complex environments require better planning (Gantt chart)

25. APST Architecture GASSIBP NFS GRAMNetSolve Condor, Ninf, Legion,.. NWS transport APIexecution API metadata API scheduler API Metadata Bookkeeper Actuator Scheduler store The Grid NetSolve Globus Legion NWS Ninf IBP transferexecutequery APST Daemon Workqueue Gantt chart heuristic algorithms Workqueue++ MinMinMaxMinSufferageXSufferage actuate report retrieve APST Client Controller triggers interacts Command-line client

26. Scheduling Results Max-min Workqueue XSufferage Sufferage Min-min Data transfer <= 40 X task computation time Scheduling event every 250 seconds Heuristics for Scheduling Parameter Sweep Applications in Grid Environments H. Casanova, A. Legrand, D. Dzagorodnov, F. Berman (HCW’00) Simulation results show that: Heuristics are worth it Xsufferage is good heuristic even when predictions are bad Complex environments require better planning