Share
Improving Bioinformatic Methods via Enhancing and Scaling Up Workflows
Workflows have proved to be an excellent medium for representing scientific methods in general and especially also in the areas life sciences and chemistry. In the last 20 years quite a few mature workflow engines and workflow editors with diverse foci and strength have been developed to support communities in managing workflows. More recent trends for enhancements of workflow systems address the usability of workflows and optimizations of their efficiency. The talk will go into detail for three methods and concepts concerned with these topics. The first topic consists of meta-workflows for quantum-chemical applications applied in the project MoSGrid (Molecular Simulation Grid) and its science gateway developed on top of WS-PGRADE. The second one addresses scaling up bioinformatic workflows with dynamic job expansion and our case study using Galaxy and Makeflow. The talk concludes with a model for balancing thread-level and task-level parallelism for data-intensive workloads on clusters and clouds on the example of aligners for next-generation sequencing data.
SciTech
Transcript
62 slides
-
Sandra Gesing Center for Research Compu6ng [email protected]
16 March 2016 Improving Bioinforma6c Methods via Enhancing and Scaling Up Workflows -
University of Notre Dame Sandra Gesing
2 hKp://chartsbin.com/view/1124 hKp://chartsbin.com/view/1124 • In the middle of nowhere of northern Indiana (1.5 h from Chicago) • 4 undergraduate colleges • ~35 research ins6tutes and centers • ~12,000 students -
Center for Research Compu6ng Sandra Gesing
4 • SoVware development and profiling • Cyberinfrastructure/science gateway development • Geographical Informa6on Systems • Visualiza6on Support • Computa6onal Scien6st support • Collabora6ve research/ grant development • System administra6on/ design and acquisi6on • ~40 researchers, research programmers, HPC specialists CRC and OIT building hKp://crc.nd.edu -
Center for Research Compu6ng Sandra Gesing
5 • Computa6onal resources: 25,000 cores+ • Storage resources: 3 PB • Visualiza6on systems • Systems for virtual hos6ng • Prototype architectures e.g., Docker, OpenStack • Access and interface to • XSEDE • Open Science Grid • Blue Waters CRC HPC Center (old Union Sta6on) -
Bioinforma6cs Sandra Gesing
6 • Genomics • Proteomics • Metabolomics • Immunomics • System biology • Molecular simula6ons • Docking • Epidemiology • … Black Swallowtail – larvae and buKerfly -
The Genomics Boom Sandra Gesing
7 February 16, 2001 biotech company Celera February 15, 2001 The Human Genome Project -
The Genomics Boom Sandra Gesing
8 Craig Venter (leV) and Francis Collins (right) -
Big Data Sandra Gesing
9 • Explosion in the quan6ty, variety and complexity of data • Ques6ons can be answered impossible to even ask about 10 years ago • Costs far reduced (e.g., Human Genome project, 15 years, ~$2 billion; today ~3 days, $1000) -
Big Data Sandra Gesing
10 hKp://www.genome.gov/images/content/cost_per_genome_oct2015.jpg -
Analysis of data Sandra Gesing
11 12181 acatttctac caacagtgga tgaggttgtt ggtctatgtt ctcaccaaat ttggtgttgt 12241 cagtctttta aattttaacc tttagagaag agtcatacag tcaatagcct tttttagctt 12301 gaccatccta atagatacac agtggtgtct cactgtgatt ttaatttgca ttttcctgct 12361 gactaattat gttgagcttg ttaccattta gacaacttca ttagagaagt gtctaatatt 12421 taggtgactt gcctgttttt ttttaattgg gatcttaatt tttttaaatt attgatttgt 12481 aggagctatt tatatattct ggatacaagt tctttatcag atacacagtt tgtgactatt 12541 ttcttataag tctgtggttt ttatattaat gtttttattg atgactgttt tttacaattg 12601 tggttaagta tacatgacat aaaacggatt atcttaacca ttttaaaatg taaaattcga 12661 tggcattaag tacatccaca atattgtgca actatcacca ctatcatact ccaaaagggc 12721 atccaatacc cattaagctg tcactcccca atctcccatt ttcccacccc tgacaatcaa 12781 taacccattt tctgtctcta tggatttgcc tgttctggat attcatatta atagaatcaa Slide copied from: Stuart Owen „Workflows with Taverna“ A sequence of connected steps in a defined order based on their control and data dependencies -
Analysis of data Sandra Gesing
12 12181 acatttctac caacagtgga tgaggttgtt ggtctatgtt ctcaccaaat ttggtgttgt 12241 cagtctttta aattttaacc tttagagaag agtcatacag tcaatagcct tttttagctt 12301 gaccatccta atagatacac agtggtgtct cactgtgatt ttaatttgca ttttcctgct 12361 gactaattat gttgagcttg ttaccattta gacaacttca ttagagaagt gtctaatatt 12421 taggtgactt gcctgttttt ttttaattgg gatcttaatt tttttaaatt attgatttgt 12481 aggagctatt tatatattct ggatacaagt tctttatcag atacacagtt tgtgactatt 12541 ttcttataag tctgtggttt ttatattaat gtttttattg atgactgttt tttacaattg 12601 tggttaagta tacatgacat aaaacggatt atcttaacca ttttaaaatg taaaattcga 12661 tggcattaag tacatccaca atattgtgca actatcacca ctatcatact ccaaaagggc 12721 atccaatacc cattaagctg tcactcccca atctcccatt ttcccacccc tgacaatcaa 12781 taacccattt tctgtctcta tggatttgcc tgttctggat attcatatta atagaatcaa Slide copied from: Stuart Owen „Workflows with Taverna“ A sequence of connected steps in a defined order based on their control and data dependencies Workflows! -
Workflows Sandra Gesing
13 • Different workflow concepts • Different workflow languages • Different workflow constructs Taverna -
Workflow Editors Sandra Gesing
14 • Different technologies (workbenches, web-‐based) • Different look-‐and-‐feel -
State of the Art Sandra Gesing
15 Data and compute-‐ intensive problems High-‐speed networks Users generally not IT specialists Tools and workflow engines Web-‐based agile frameworks Distributed data and compu6ng infrastructures -
Challenge for Developers Sandra Gesing
16 Data and compute-‐ intensive problems High-‐speed networks Tools and workflow engines Web-‐based agile frameworks Distributed data and compu6ng infrastructures Users generally not IT specialists Need for intui6ve and efficient workflows! -
Challenge for Developers Sandra Gesing
17 Data and compute-‐ intensive problems High-‐speed networks Tools and workflow engines Web-‐based agile frameworks Distributed data and compu6ng infrastructures Users generally not IT specialists -
Usability Sandra Gesing
19 “AVer all, usability really just means that making sure that something works well: that a person … can use the thing -‐ whether it's a Web site, a fighter jet, or a revolving door -‐ for its intended purpose without geung hopelessly frustrated.” (Steve Krug in “Don't make me think!: A Common Sense Approach to Web Usability”, 2005) -
Usability Sandra Gesing
20 ... and users should get more features easily... -
Workflow Enhancements Sandra Gesing
21 • Logical level: Meta-‐workflows Herres-‐Pawlis, S., Hoffmann, A., Rösener, T., Krüger, J., Grunzke, R., and Gesing, S. “Mul6-‐layer Meta-‐ metaworkflows for the Evalua6on of Solvent and Dispersion Effects in Transi6on Metal Systems Using the MoSGrid Science Gateways”Science Gateways (IWSG), 2015 7th Interna6onal Workshop on, pp.47-‐52, 3-‐5 June 2015, IEEE Xplore, doi: 10.1109/IWSG.2015.13 • System level: Combina6on of strengths of workflow systems Hazekamp, N., Sarro, J., Choudhury, O., Gesing, S., ScoK Emrich and Thain, D. “Scaling Up Bioinforma6cs Workflows with Dynamic Job Expansion: A Case Study Using Galaxy and Makeflow”, e-‐Science (e-‐Science), 2015 IEEE 11th Interna6onal Conference on, pp.332-‐341, Aug. 31 2015-‐Sept. 4 2015 • Predic6on: Model for op6miza6on of tasks and threads Choudhury, O., Rajan, D., Hazekamp, N., Gesing, S., Thain, D., and Emrich, S. “Balancing Thread-‐level and Task-‐ level Parallelism for Data-‐Intensive Workloads on Clusters and Clouds”, Cluster Compu6ng (CLUSTER), 2015 IEEE Interna6onal Conference on, pp.390-‐393, 8-‐11 Sept. 2015, doi:10.1109/CLUSTER.2015.60 -
Molecular Simula6ons Sandra Gesing
22 • Predic6on and analysis of molecular structures • Numerous applica6ons, e.g. • Materials science • Drug design ligands target docking ? -
Molecular Simula6ons ligands target docking binding
energy scoring func6ons binding pocket • Predic6on and analysis of molecular structure • Numerous applica6ons, e.g. • Materials science • Drug design Sandra Gesing 23 -
MoSGrid Science Gateway Molecular Simula6on Grid •
Science gateway integrated with underlying compute and data management infrastructure • Distributed workflow management • Data repository • Metadata management Sandra Gesing 24 -
MoSGrid Science Gateway User Interface WS-‐PGRADE Liferay
DCI Resources Middleware Layer UNICORE XtreemFS High-‐Level Middleware Service Layer gUSE Sandra Gesing 25 -
MoSGrid Science Gateway • Three domains (QC, MD, Docking)
• Gaussian, NWCHem, Gromacs, CADDSuite, … • Compute and data intensive • embarrassingly parallel, MPI and GPUs • Shared workflows, metaworkflows, data and knowledge • Distributed compu6ng and data management via UNICORE and XtreemFS Sandra Gesing 26 -
Quantum Chemistry
Sandra Gesing 27 -
Quantum Chemistry
Sandra Gesing 28 -
Molecular Dynamics
Sandra Gesing 29 -
Docking
Sandra Gesing 30 -
Visualiza6on
Tes6ng of ChemDoodle and MolCAD web.chemdoodle.com molcad.de Sandra Gesing 31 -
Scaling Up Workflows Sandra Gesing
37 # Machines # Cores Data ParBBoning Save -
Scaling Up Workflows Sandra Gesing
38 Galaxy -
Scaling Up Workflows Sandra Gesing
39 Simple Workflow in Galaxy Problem: As Size increases so does Time -
Scaling Up Workflows Sandra Gesing
40 Workflow with Parallelism added in Galaxy Problem: Tools must be updated every change in Parallelism/Relies on Scien6st -
Scaling Up Workflows Sandra Gesing
41 Workflow Dynamically Expanded behind Galaxy -
Scaling Up Workflows Sandra Gesing
44 Makeflow • Task Structure INPUTS : OUTPUTS COMMAND • Directed Acyclic Graph (DAG) • Programma6cally Generated -
Scaling Up Workflows Sandra Gesing
47 Job Sandbox – Log file crea6on for cleanup -
Scaling Up Workflows Sandra Gesing
49 Using Dynamic Job Expansion we were able to scale up a workflow without requiring the huge amount of 6me to process it Found viable solu6ons for: • Using Work Queue we u6lized 100s of cores from a Condor Pool • Cleaning Sandbox using knowledge of intermediates and logging • Explored methods to transmit needed environments such as executables and Java 61.5X speed-‐up on 32 GB dataset u6lizing these methods -
Thread-‐level and Task-‐level Parallelism Sandra Gesing
50 • Develop predictive performance models for an application domain • Achieve acceptable performance the first time • Optimize resource utilization • Execution time • Memory usage -
Thread-‐level and Task-‐level Parallelism • WorkQueue master-worker framework
• Sun Grid Engine (SGE) batch system Sandra Gesing 51 -
Thread-‐level and Task-‐level Parallelism Sandra Gesing
52 1. Applica6on-‐level model for 6me: (,,)= 1/ + 2 2. Applica6on-‐level model for memory: (,)= γ1R +γ2N 3. System-‐level model for 6me: =1/ +2(/ +/ )+3T(R,/ ,)∗/ + 4/ +5/ 4. System-‐level model for memory: (,)=ϕ1R +ϕ2Q -
Thread-‐level and Task-‐level Parallelism Sandra Gesing
53 7 data points (R) 7 data points (Q) 7 data points (N) 343 data points Data CollecBon Training data Regression Model Training Accuracy Test MAPE TesBng Regression Coefficient s Tes6n g data -
Thread-‐level and Task-‐level Parallelism Sandra Gesing
54 Avg. MAPE = 3.1 MAPE = Mean Absolute Percentage Error -
Thread-‐level and Task-‐level Parallelism Sandra Gesing
55 For the given dataset, K* = 90, N* = 4 -
Result Sandra Gesing
56 # Cores/ Task # Tasks Predicted Time (min) Speedup Estimated EC2 Cost ($) Estimated Azure Cost ($) 1 360 70 6.6 50.4 64.8 2 180 38 12.3 25.2 32.4 4 90 24 19.5 18.9 32.4 8 45 27 17.3 18.9 32.4 -
Challenges – Novel and Old... Sandra
Gesing Science Gateways 57 … require novel solu6ons! -
EU COST Ac6on cHiPSet (IC1406) Sandra Gesing
Science Gateways 58 cHiPSet – High Performance Modeling and Simula6on for Big Data Applica6ons • April 2015 – April 2019 • 15 countries -‐ 12 COST, 3 non-‐COST (US, China, Australia) • 37 reseach organiza6ons/companies (31 COST, 6 non-‐ COST) hKp://www.cost.eu/COST_Ac6ons/ict/Ac6ons/IC1406 -
EU COST Ac6on cHiPSet Sandra Gesing
Science Gateways 59 -
cHiPSet -‐ Collabora6ons Sandra Gesing
Science Gateways 60 Partner Projects • NESUS (Network for Sustainable Ultrascale Compu6ng) hKp://www.nesus.eu/ • MPM4CPS (Mul6-‐Paradigm Modelling for Cyber-‐ Physical Systems) hKp://www.mpm4cps.eu/ And maybe YOU? -
Acknowledgements Sandra Gesing
Science Gateways 61 Logical level • Richard Grunzke • Sonja Herres-‐Pawlis • Alexander Hoffmann • Jens Krüger • MTA SZTAKI • University of Westminster System level and predic6on • Nicholas Hazekamp • Olivia Choudhury • Douglas Thain • Scott Emrich • Notre Dame Bioinformatics Lab • The Cooperative Computing Lab, University of Notre Dame