Qiu bosc2010

1. Cloud Technologies and Their ApplicationsThe Bioinformatics Open Source Conference (BOSC 2010) Boston, MassachusettsJudy Qiuhttp://salsahpc.indiana.edu Assistant Director, Pervasive Technology InstituteAssistant Professor, School of Informatics and ComputingIndiana University

2. Data Explosion and ChallengesData DelugeCloud TechnologiesWhy ?How ?Life Science ApplicationsParallel ComputingWhat ?

3. Data We’re Looking atPublic Health Data (IU Medical School & IUPUI Polis Center)(65535 Patient/GIS records / 54 dimensions each)Biology DNA sequence alignments (IU Medical School & CGB) (10 million Sequences / at least 300 to 400 base pair each)NIH PubChem (IU Cheminformatics) (60 million chemical compounds/166 fingerprints each)High volume and high dimension require new efficient computing approaches!

4. Some Life Sciences ApplicationsEST (Expressed Sequence Tag)sequence assembly program using DNA sequence assembly program software CAP3.Metagenomics and Alu repetition alignment using Smith Waterman dissimilarity computations followed by MPI applications for Clustering and MDS (Multi Dimensional Scaling) for dimension reduction before visualizationMapping the 60 million entries in PubCheminto two or three dimensions to aid selection of related chemicals with convenient Google Earth like Browser. This uses either hierarchical MDS (which cannot be applied directly as O(N2)) or GTM (Generative Topographic Mapping).Correlating Childhood obesity with environmental factorsby combining medical records with Geographical Information data with over 100 attributes using correlation computation, MDS and genetic algorithms for choosing optimal environmental factors.

5. DNA Sequencing PipelineMapReduceIllumina/Solexa Roche/454 Life Sciences Applied Biosystems/SOLiDPairwiseclusteringBlocking MDSMPIModern Commerical Gene SequencesVisualizationPlotvizSequencealignmentDissimilarityMatrixN(N-1)/2 valuesblockPairingsFASTA FileN SequencesRead AlignmentThis chart illustrate our research of a pipeline mode to provide services on demand (Software as a Service SaaS)

6. User submit their jobs to the pipeline. The components are services and so is the whole pipeline.Internet

7. Cloud Services and MapReduceCloud TechnologiesData DelugeLife ScienceApplicationsParallel Computing

8. Clouds as Cost Effective Data Centers7Builds giant data centers with 100,000’s of computers; ~ 200-1000 to a shipping container with Internet access “Microsoft will cram between 150 and 220 shipping containers filled with data center gear into a new 500,000 square foot Chicago facility. This move marks the most significant, public use of the shipping container systems popularized by the likes of Sun Microsystems and Rackable Systems to date.”―News Release from Web

9. Clouds hide Complexity8CyberinfrastructureIs “Research as a Service”SaaS: Software as a Service(e.g. Clustering is a service)PaaS: Platform as a ServiceIaaS plus core software capabilities on which you build SaaS(e.g. Azure is a PaaS; MapReduce is a Platform)IaaS(HaaS): Infrasturcture as a Service (get computer time with a credit card and with a Web interface like EC2)

10. Commercial CloudSoftware

11. MapReduceMap(Key, Value) Reduce(Key, List<Value>) A parallel Runtime coming from Information RetrievalData PartitionsA hash function maps the results of the map tasks to r reduce tasksReduce OutputsImplementations support:Splitting of dataPassing the output of map functions to reduce functionsSorting the inputs to the reduce function based on the intermediate keysQuality of services

12. Edge : communication pathVertex :execution task Hadoop & DryadLINQApache HadoopMicrosoft DryadLINQStandard LINQ operationsData/Compute NodesMaster NodeDryadLINQ operationsJobTrackerMMMMRRRRHDFSNameNodeDatablocks12DryadLINQ Compiler2334Directed Acyclic Graph (DAG) based execution flowsDryad process the DAG executing vertices on compute clustersLINQ provides a query interface for structured dataProvide Hash, Range, and Round-Robin partition patterns Apache Implementation of Google’s MapReduceHadoop Distributed File System (HDFS) manage dataMap/Reduce tasks are scheduled based on data locality in HDFS (replicated data blocks)Dryad Execution EngineJob creation; Resource management; Fault tolerance& re-execution of failed taskes/vertices

13. Applications using Dryad & DryadLINQInput files (FASTA)CAP3 - Expressed Sequence Tag assembly to re-construct full-length mRNACAP3CAP3CAP3DryadLINQOutput filesPerform using DryadLINQ and Apache Hadoop implementationsSingle “Select” operation in DryadLINQ“Map only” operation in Hadoop X. Huang, A. Madan, “CAP3: A DNA Sequence Assembly Program,” Genome Research, vol. 9, no. 9, pp. 868-877, 1999.

14. Classic Cloud ArchitectureAmazon EC2 and Microsoft AzureMapReduce ArchitectureApache Hadoop and Microsoft DryadLINQHDFSInput Data SetData FileMap()Map()ExecutableOptionalReducePhaseReduceResultsHDFS

15. Usability and Performance of Different Cloud ApproachesCap3 PerformanceCap3 EfficiencyEfficiency = absolute sequential run time / (number of cores * parallel run time)

16. Hadoop, DryadLINQ - 32 nodes (256 cores IDataPlex)

17. EC2 - 16 High CPU extra large instances (128 cores)

18. Azure- 128 small instances (128 cores)

19. Ease of Use – Dryad/Hadoop are easier than EC2/Azure as higher level models

20. Lines of code including file copyAzure : ~300 Hadoop: ~400 Dyrad: ~450 EC2 : ~700

21. Table 1 : Selected EC2 Instance Types

22. 4096 Cap3 data files : 1.06 GB / 1875968 reads (458 readsX4096)..Following is the cost to process 4096 CAP3 files..Amortized cost in Tempest (24 core X 32 nodes, 48 GB per node) = 9.43$(Assume 70% utilization, write off over 3 years, include support)

23. Data Intensive ApplicationsCloud TechnologiesData DelugeLife Science ApplicationsParallel Computing

24. Alu and Metagenomics Workflow“All pairs” problem Data is a collection of N sequences. Need to calcuate N2dissimilarities (distances) between sequnces (all pairs).These cannot be thought of as vectors because there are missing characters

25. “Multiple Sequence Alignment” (creating vectors of characters) doesn’t seem to work if N larger than O(100), where 100’s of characters long.Step 1: Can calculate N2 dissimilarities (distances) between sequencesStep 2: Find families by clustering (using much better methods than Kmeans). As no vectors, use vector free O(N2) methodsStep 3: Map to 3D for visualization using Multidimensional Scaling (MDS) – also O(N2)Results: N = 50,000 runs in 10 hours (the complete pipeline above) on 768 coresDiscussions:Need to address millions of sequences …..Currently using a mix of MapReduce and MPITwister will do all steps as MDS, Clustering just need MPI Broadcast/Reduce

26. All-Pairs Using DryadLINQ125 million distances4 hours & 46 minutesCalculate Pairwise Distances (Smith Waterman Gotoh)Calculate pairwise distances for a collection of genes (used for clustering, MDS)Fine grained tasks in MPICoarse grained tasks in DryadLINQPerformed on 768 cores (Tempest Cluster)Moretti, C., Bui, H., Hollingsworth, K., Rich, B., Flynn, P., & Thain, D. (2009). All-Pairs: An Abstraction for Data Intensive Computing on Campus Grids. IEEE Transactions on Parallel and Distributed Systems, 21, 21-36.

27. Biology MDS and Clustering ResultsAlu FamiliesThis visualizes results of Alu repeats from Chimpanzee and Human Genomes. Young families (green, yellow) are seen as tight clusters. This is projection of MDS dimension reduction to 3D of 35399 repeats – each with about 400 base pairsMetagenomicsThis visualizes results of dimension reduction to 3D of 30000 gene sequences from an environmental sample. The many different genes are classified by clustering algorithm and visualized by MDS dimension reduction

28. Hadoop/Dryad ComparisonInhomogeneous Data IInhomogeneity of data does not have a significant effect when the sequence lengths are randomly distributedDryad with Windows HPCS compared to Hadoop with Linux RHEL on Idataplex (32 nodes)

29. Hadoop/Dryad ComparisonInhomogeneous Data IIThis shows the natural load balancing of Hadoop MR dynamic task assignment using a global pipe line in contrast to the DryadLinq static assignmentDryad with Windows HPCS compared to Hadoop with Linux RHEL on Idataplex (32 nodes)

30. Hadoop VM Performance DegradationPerf. Degradation = (Tvm – Tbaremetal)/Tbaremetal15.3% Degradation at largest data set size

31. Parallel Computing and SoftwareCloud TechnologiesData DelugeLife Science ApplicationsParallel Computing

32. Twister(MapReduce++)Pub/Sub Broker NetworkMap WorkerStreaming based communication

33. Intermediate results are directly transferred from the map tasks to the reduce tasks – eliminates local files

34. Cacheablemap/reduce tasks

35. Static data remains in memory

36. Combine phase to combine reductions

37. User Program is the composer of MapReduce computations

38. Extendsthe MapReduce model to iterativecomputationsMStaticdataConfigure()Worker NodesReduce WorkerRDDMRDriverUserProgramIterateMRDeamonDMMMMData Read/WriteRRRRUser Programδ flowCommunicationMap(Key, Value) File SystemData SplitReduce (Key, List<Value>) Close()Combine (Key, List<Value>)Different synchronization and intercommunication mechanisms used by the parallel runtimes

39. Twister New Release

40. Iterative ComputationsK-meansMatrix MultiplicationPerformance of K-Means Parallel Overhead Matrix Multiplication

41. Dimension Reduction AlgorithmsMultidimensional Scaling (MDS) [1]Given the proximity information among points.

42. Optimization problem to find mapping in target dimension of the given data based on pairwise proximity information while minimize the objective function.

43. Objective functions: STRESS (1) or SSTRESS (2)

44. Only needs pairwise distances ijbetween original points (typically not Euclidean)

45. dij(X) is Euclidean distance between mapped (3D) pointsGenerative Topographic Mapping (GTM) [2]Find optimal K-representations for the given data (in 3D), known as K-cluster problem (NP-hard)

46. Original algorithm use EM method for optimization

47. Deterministic Annealing algorithm can be used for finding a global solution

48. Objective functions is to maximize log-likelihood:[1] I. Borg and P. J. Groenen. Modern Multidimensional Scaling: Theory and Applications. Springer, New York, NY, U.S.A., 2005.[2] C. Bishop, M. Svens´en, and C. Williams. GTM: The generative topographic mapping. Neural computation, 10(1):215–234, 1998.

49. Science Cloud (Dynamic Virtual Cluster) ArchitectureSmith Waterman Dissimilarities, CAP-3 Gene Assembly, PhyloD Using DryadLINQ, High Energy Physics, Clustering, Multidimensional Scaling, Generative Topological MappingApplicationsServices and WorkflowMicrosoft DryadLINQ / MPIApache Hadoop / Twister/ MPIRuntimesLinux Bare-systemWindows Server 2008 HPCBare-systemLinux Virtual MachinesWindows Server 2008 HPCInfrastructure softwareXen VirtualizationXen VirtualizationXCAT InfrastructureHardwareiDataplex Bare-metal NodesDynamic Virtual Cluster provisioning via XCATSupports both stateful and stateless OS images

50. Dynamic Virtual ClustersMonitoring & Control InfrastructureMonitoring InterfaceMonitoring InfrastructureDynamic Cluster ArchitecturePub/Sub Broker NetworkSW-G Using Hadoop SW-G Using Hadoop SW-G Using DryadLINQVirtual/Physical ClustersLinux Bare-systemLinux on XenWindows Server 2008 Bare-systemSwitchable clusters on the same hardware (~5 minutes between different OS such as Linux+Xen to Windows+HPCS)Support for virtual clustersSW-G : Smith Waterman Gotoh Dissimilarity Computation as an pleasingly parallel problem suitable for MapReduce style applicationsXCAT InfrastructureSummarizeriDataplex Bare-metal Nodes (32 nodes)XCAT InfrastructureSwitcheriDataplex Bare-metal Nodes

51. SALSA HPC Dynamic Virtual Clusters DemoAt top, these 3 clusters are switching applications on fixed environment. Takes ~30 Seconds.

52. At bottom, this cluster is switching between Environments – Linux; Linux +Xen; Windows + HPCS. Takes about ~7 minutes.

53. It demonstrates the concept of Science on Clouds using a FutureGrid cluster.FutureGrid: a Grid Testbed IU Cray operational, IU IBM (iDataPlex) completed stability test May 6 UCSD IBM operational, UF IBM stability test completes ~ May 12Network, NID and PU HTC system operationalUC IBM stability test completes ~ May 27; TACC Dell awaiting delivery of componentsNID: Network Impairment DevicePrivatePublicFG Network

54. Summary of Initial ResultsCloud technologies (Dryad/Hadoop/Azure/EC2) promising for Biology computationsDynamic Virtual Clusters allow one to switch between different modesOverhead of VM’s on Hadoop (15%) acceptableInhomogeneous problems currently favors Hadoop over DryadTwister allows iterative problems (classic linear algebra/datamining) to use MapReduce model efficientlyPrototype Twister released

56. ReferencesTwister  Open Source Iterative MapReduce Softwarewww.iterativemapreduce.orgSALSA Project salsahpc.indiana.eduFutureGrid Projectfuturegrid.orgSponsorsMicrosoft, NIH, NSF, Pervasive Technology Institute

57. MapReduce and Clouds for Science http://salsahpc.indiana.eduIndiana University BloomingtonJudy Qiu, SALSA GroupSALSA project (salsahpc.indiana.edu) investigates new programming models of parallel multicore computing and Cloud/Grid computing. It aims at developing and applying parallel and distributed Cyberinfrastructure to support large scale data analysis. We illustrate this with a study of usability and performance of different Cloud approaches. We will develop MapReduce technology for Azure that matches that available on FutureGrid in three stages: AzureMapReduce (where we already have a prototype), AzureTwister, and TwisterMPIReduce. These offer basic MapReduce, iterative MapReduce, and a library mapping a subset of MPI to Twister. They are matched by a set of applications that test the increasing sophistication of the environment and run on Azure, FutureGrid, or in a workflow linking them.Iterative MapReduce using Java Twisterhttp://www.iterativemapreduce.org/Twister supports iterative MapReduce Computations and allows MapReduce to achieve higher performance, perform faster data transfers, and reduce the time it takes to process vast sets of data for data mining and machine learning applications. Open source code supports streaming communication and long running processes. MPI is not generally suitable for clouds. But the subclass of MPI style operations supported by Twister – namely, the equivalent of MPI-Reduce, MPI-Broadcast (multicast), and MPI-Barrier – have large messages and offer the possibility of reasonable cloud performance. This hypothesis is supported by our comparison of JavaTwister with MPI and Hadoop. Many linear algebra and data mining algorithms need only this MPI subset, and we have used this in our initial choice of evaluating applications. We wish to compare Twister implementations on Azure with MPI implementations (running as a distributed workflow) on FutureGrid. Thus, we introduce a new runtime, TwisterMPIReduce, as a software library on top of Twister, which will map applications using the broadcast/reduce subset of MPI to Twister.Architecture of TwisterMapReduce on Azure − AzureMapReduceAzureMapReduce uses Azure Queues for map/reduce task scheduling, Azure Tables for metadata and monitoring data storage, Azure Blob Storage for input/output/intermediate data storage, and Azure Compute worker roles to perform the computations. The map/reduce tasks of the AzureMapReduce runtime are dynamically scheduled using a global queue.Usability and Performance of Different Cloud and MapReduce ModelsThe cost effectiveness of cloud data centers combined with the comparable performance reported here suggests that loosely coupled science applications will increasingly be implemented on clouds and that using MapReduce will offer convenient user interfaces with little overhead. We present three typical results with two applications (PageRank and SW-G for biological local pairwise sequence alignment) to evaluate performance and scalability of Twister and AzureMapReduce. Architecture of AzureMapReduceArchitecture of TwisterMPIReduceParallel Efficiency of the different parallel runtimes for the Smith Waterman Gotoh algorithmTotal running time for 20 iterations of Pagerank algorithm on ClueWeb data with Twister and Hadoop on 256 coresPerformance of AzureMapReduce on Smith Waterman Gotoh distance computation as a function of number of instances used

Qiu bosc2010

More Related Content

What's hot (18)

Viewers also liked (20)

Similar to Qiu bosc2010 (20)

More from BOSC 2010 (20)

Recently uploaded (20)

Qiu bosc2010

Editor's Notes