Research at the

1. Research atthe MartaKwiatkowska School of Computer Science www.mesc.bham.ac.uk

2. Overview • The Midlands e-Science Centre • Area of ExcellenceModelling and Analysis of Large Complex Systems • Applications focus, rather than Grid middleware • Infrastructure • Computing and collaborative working • Centres • Projects • e-Science and Grid collaborations • Parallelising probabilistic model checking – a glimpse • Summary

3. Complex systems New field of science - study how parts of a system give rise to the collective behaviours, and how it interacts with its environment. Social science, medicine, weather, engineering, economy, management...

4. The Midlands e-Science Centre • Virtual Centre • collaborative • University of Birmingham • home Computer Science • Physics and Astronomy • Chemical Sciences • Biosciences • Engineering • Geography, Earth and Env. Sci. • Mathematics and Statistics • Medical School • Information Services • University of Warwick • Centre for Scientific Computing • University of Coventry • University of Wolverhampton

5. Infrastructure and Facilities • Computing and collaborative working • 1000 CPU cluster in planning phase • Dedicated 100 CPU e-Science centre cluster • Various clusters on campus • Access Grid Node rooms (x2 at Birmingham +2) • Centres • VISTA, state-of-the-art HP visualisation centre • CERCIA, research into computational intelligence • CETADL, educational technology and distance learning • Henry Wellcome NMR • Informatics Collaborative Research Network • Collaborations within Birmingham

6. Visual and Spatial Technology Centre • Set up in partnership with HP • £4M investment • Association with several industrial partners (AVS, CFX, Fakespace, etc) • Scientific visualisation • geodata, medical imaging • Information visualisation • knowledge discovery • Data representation • understanding complex data • Immersive environments Part of the internal structure of a hydrogen atom. Image fusion of a series of MRI scans. www.vista.bham.ac.uk/index.htm

7. Research at MeSC • Research themes • Simulation of evolving systems of interacting components • Large-scale Grid-enabled distributedsimulation • Mathematical solutions of large complex systems • Data mining and large-scale visualisation • Examples of expertise • bioinformatics, genomics, proteomics, etc • healthcare informatics • computational modelling of complex systems: biological processes, engineering structures, etc • security, reliability/correctness assurance • etc www.mesc.bham.ac.uk

8. Projects • At Birmingham • Biological modelling • IntegrativeBiology, computational model of human heart, cancer modelling • Predictive modelling, using probabilistic model checking with PRISM • Medical informatics • cancerGrid, Cancer clinical trials: entry, randomisation and follow-up into hospital care • e-TUMOUR (EU FP6 IP), MRI, web-enabled diagnosis, prognosis • ePCRN (Primary Care), randomised trials • Physics • GridPP, LIGO & LISA (GW), STAR (Nuclear Physics)

9. Projects continued… • At Birmingham • Engineering • Fluidised beds, particulate processes (Engineering) • Simulation • DS-Grid, Grid-enabled distributed simulation, applications in logistics, business • Social sciences • COSI (Complexity in Social Sciences, EU) • At Warwick • PACE, Performance Analysis and Characterisation Environment • Molecular modelling

10. e-Science pilot project, funded by EPSRC, £2.4m • Two components, computational modelling of human heart and modelling of cancer • Oxford, Nottingham, UCL, Leeds, Birmingham, Sheffield • Cancer modelling • 3 institutions, 2 RFs & 10 PhDs • Multiscale • Tissue, cellular, subcellular, drugresponse • Colorectal cancer • Crypt formation • Cell migration in crypt, role of stem cells

11. MRC funded e-Science project £2.2m • a consortium of specialists in oncology and software engineering drawn from UK Universities • Cambridge, Oxford, UCL, Birmingham, Belfast • Modular, distributed software solutions for the key problems of clinical cancer informatics • clinical trial patient entry, • randomisation and follow-up; • storage and analysis of complex datasets; • linking trial and epidemiology data with profiling information

12. Fluidised beds • Modelling and Simulation (DEM) of Particulate Processes • discontinuous, composed of manymillions of particles • particles interact in various ways • aim to calculate properties of substance: elasticity, texture, feel • Grid technology needed because of sheer scale of models • Research in Chemical and Civil Engineering • funding grid enable from EPSRC Pharmaceuticals, foods, powders, aerosols, soils, ... www.eng.bham.ac.uk/chemical/

13. DS-Grid: distributed simulation • University of Birmingham, Nottingham and Nanyang Technological University, Singapore • One of four “Sister Projects” funded by the e-Science Core Programme - first UK-Singapore Grid link • A Grid “plug-and-play” distributed collaborative simulation environment

14. Grid-enabled PRISM • Probabilistic model checking • What does it involve? • Symbolic MTBDD-based techniques • The PRISM tool • Parallel symbolic CTMC solution • Why Gauss-Seidel? • Wavefront technology • Implementation • Experimental results • Applications • Analysis of Internet protocols • Modelling of biological processes, sponsored by Microsoft

15. 0.4 0.3 ü Probabilistic model or û or The probability send  P¸ 0.9(deliver) State 5: 0.6789 State 6:0.9789 State 7:1.0 … State 12: 0 State 13: 0.1245 Probabilistic temporal logic specification Probabilistic model checking… in a nutshell Probabilistic Model Checker

16. The PRISM probabilistic model checker • History • Implemented at the University of Birmingham • First public release September 2001, ~7 years development • Open source, GPL licence, available freely • >2500 downloads, many users worldwide, >80 papers using PRISM, ~50 case studies and 6 flaws, connection to several tools, … • Approach • Based on symbolic, BDD-based techniques • Multi-Terminal BDDs, first algorithm [ICALP’97] • www.cs.bham.ac.uk/~dxp/prism/ • For software, publications, case studies, taught courses using PRISM, etc

17. Screenshot of PRISM GUI

18. Overview of PRISM • Functionality • Implements temporal logic probabilistic model checking • Construction of models: discrete and continuous Markov chains (DTMCs/CTMCs), and Markov decision processes (MDPs) • Probabilistic temporal logics:PCTL and CSL • Extension with costs/rewards, expectation operator • Underlying computation combines graph-theoretical algorithms • Reachability, qualitative model checking, BDD-based with numerical computation – iterative methods • Linear equation system solution - Jacobi, Gauss-Seidel, ... • Uniformisation (CTMCs) • Dynamic programming (MDPs) • Explicit and symbolic (MTBDDs, etc.) • Now Grid enabled, see [AHM’05]

19. PRISM real-world case studies • CTMCs • Dynamic power management [HLDVT’02, FAC 2005] • Dependability of embedded controller [INCOM’04] • RKIP-inhibited ERK pathway (by Calder et al) [CMSB’01] • Eukaryotic cell cycle control (based on Lecca & Priami) • MDPs/DTMCs • Bluetooth device discovery [ISOLA’04] • Crowds anonymity protocol (by Shmatikov) [CSFW’02, JSC 2003] • Randomised consensus [CAV’01,FORTE’02] • Contract signing protocols (by Norman & Shmatikov) [FASEC’02] • Reliability of NAND multiplexing (with Shukla) [VLSI’04,TCAD 2005] • PTAs • IPv4 ZeroConf dynamic configuration [FORMATS’03] • Root contention in IEEE 1394 FireWire [FAC 2003, STTT 2004] • IEEE 802.11 (WiFi) Wireless LAN MAC protocol [PROBMIV’02]

20. Why Grid-enable? • Experience with PRISM indicates • Symbolic representation very compact, >1010 states for CTMCs • Extremely large models feasible depending on regularity • Numerical computation often slow • Sampling-based computation can be even slower • Can parallelise the symbolic approach • Facilitates extraction of the dependency information • Compactness enables storage of the full matrix at each node • Focus on steady-state solution for CTMCs, can generalise • Use wavefront techniques to parallelise • Easy to distribute sampling-based computation • Individual simulation runs are independent

21. Numerical Solution for CTMCs • Steady-state probability distribution can be obtained by solving linear equation system: • pQ = 0 with constraint Sipi = 1 • We consider the more general problem of solving: • Ax = b, • where A is nn matrix, b vector of length n • Numerical solution techniques • Direct, not feasible for very large models • Iterative stationary(Jacobi, Gauss-Seidel), memory efficient • Projection methods (Krylov, CGS, …), fastest convergence, but require several vectors • Transient probabilities similar • Computed via an iterative method (uniformisation)

22. Gauss-Seidel • Computes one matrix row at a time • Updates ith element using most up-to-date values • Computation for a single iteration, nn matrix: • for (0 · i · n-1) • xi := (bi - å0· j· n-1, j¹ i Aij¢ xj) / Aii • Can be reformulated in block form, NN blocks, length M • for (0 · p · N-1) • v := b(p) • for each block A(pq)withq¹p • v := v - A(pq) x(q) • for(0· i· M-1,i¹j) • x(p)i:= (vi-S0· j· M A(pp)ij¢ x(p)j ) / A(pp)ii • Computes one matrix block at a time

23. Parallelising Gauss-Seidel • Inherently sequential for dense matrices • Uses results from current and previous iterations • Permutation has no effect on correctness of the result • Can be exploited to achieve parallelisation for certain sparse matrix problems, e.g. Koester, Ranka & Fox 1994 • The block formulation helps, although • Requires row-wise access to blocks and block entries • Need to respect computational dependencies, i.e. when computing vector block x(p), use values from current iteration for blocks q < p, and from previous iteration for q > p • Idea: propose to use wavefront techniques • Extract dependency information and form execution schedule

24. Symbolic techniques for CTMCs • Explicit matrix representation • Intractable for very large matrices • Symbolic representations • e.g. Multi-Terminal Binary Decision Diagrams (MTBDDs), matrix diagrams and Kronecker representation • Exploit regularity to obtain compact matrix storage • Also faster model construction, reachability, etc • Sometimes also beneficial for vector storage • This paper uses MTBDDs (Clarke et al) and derived structures • Underlying data structure of the PRISM model checker • Enhanced with caching-based techniques that substantially improve numerical efficiency

25. MTBDD data structures • Recursive, based on Binary Decision Diagrams (BDDs) • Stored in reduced form (DAG), with isomorphic subtrees stored only once • Exploit regularity to obtain compact matrix storage

26. Matrices as MTBDDs • Representation • Root represents the whole matrix • Leaves store matrix entries, reachable by following paths from the root node

27. Matrices as MTBDDs • Recursively descending through the tree • Divides the matrix into submatrices • One level, divide into two submatrices

28. Matrices as MTBDDs • Recursively descending through the tree • Provides a convenient block decomposition • Two levels, divide into four blocks

29. A two-layer structure from MTBDDs • Can obtain block decomposition, store as two sparse matrices • Enables fast row-wise access to blocks and block entries (Parker’02, Mehmood’05)

30. Wavefront techniques • An approach to parallel programming, e.g. Joubert et al ‘98 • Divide a computation into many tasks • Form a schedule for these tasks • A schedule contains several wavefronts • Eachwavefrontcomprises tasks that arealgorithmically independentof each other • i.e. correctness is not affected by the order of execution • The execution is carried out from one wavefront to another • Tasks assigned according to the dependency structure • Each wavefront contains tasks that can be executed in parallel

31. Dependency graph from MTBDD • By traversal of top levels of MTBDD, as for top layer

32. Generating a Wavefront Schedule • By colouring the dependency graph • Can generate a schedule to let the computation perform from one colour to another

33. A Wavefront Parallel Gauss-Seidel • The computation ordering of Gauss-Seidel method is determined by the colouring scheme • Each processor is responsible for the computation and storage of every vector block associated with tasks assigned to it • The computation progresses from onecolour to another • Tasks with the same colour are algorithmically independentof each other and can be executed in parallel • The completion of a task may enable tasks in the next colour to be executed • Note • Can store full matrix at each node, hence reduced communication costs, since only vector blocks need to be exchanged • (With two-layer data structure), faster access to blocks and block entries than conventional MTBDDs • Number of iterations can vary between runs

34. Implementation • Implemented on a Ethernet and Myrinet-enabled PC cluster • Use MPI (the MPICH implementation) • Prototype extension for PRISM, uses PRISM numerical engines and CUDD package for MTBDDs (Somenzi) • 32 nodes available • Evaluated on a range of benchmarks • Kanban, FMS and Polling system • Optimisations for PC-cluster environments • Non-blocking inter-processor communication is used to interleave communication and computation • Load-balancing is implemented to distribute computation load evenly between processors and minimise communication load • Cache mechanism is used to reduce communication further

35. Experimental results: models • Parameters and statistics of models • Include Kanban 9,10 and FMS 13, previously intractable • All compact, requiring less than 1GB

36. Experimental results: time • Total execution times (in seconds) with 1 to 32 nodes • Termination condition maximum relative difference 10-6 • Block numbers selected to minimise storage

37. Experimental results: FMS speed-up

38. Experimental results: Kanban speed-up

39. Summary • Our parallelisation of Gauss-Seidel • Allows much larger CTMC models to be solved • Has good overall speedup • Symbolic approach particularly well suited to Wavefront parallelisation of Gauss-Seidel • Easy to extract task dependency information • Reduced memory requirement and communication load • Gauss-Seidel excellent candidate to solve very large linear equation systems • Small memory requirement (only requires one iteration vector, vs 2 for Jacobi and 6 for CGS) • Method generalises to other symbolic techniques and application domains

40. J. Rutten, M. Kwiatkowska, G. Norman and D. ParkerMathematical Techniques for Analyzing Concurrent and Probabilistic SystemsP. Panangaden and F. van Breugel (editors), CRM Monograph Series, vol. 23, AMSMarch 2004 For more information… www.cs.bham.ac.uk/~dxp/prism/ • Case studies, statistics, group publications • Download, version 2.0 (approx. 2200 users) • Linux/Unix, Windows, Macintosh versions • Publications by others and courses that feature PRISM…