1 Data Intensive Science Education Thomas J. Hacker Associate Professor, Computer & Information Technology Purdue University, West Lafayette, Indiana USA Gjesteprofessor (Visiting Professor), Department of Electrical Engineering and Computer Science University of Stavanger, Norway EU-China-Nord America Workshop on HPC Cloud and Big Data June 20, 2013 University of Stavanger, Norway
2 Introduction and Motivation Theory and Experiment (1800s) Computational Simulation Third leg of science Past 50 years or so (1950s) Data (21st century science) Fourth leg of science Researchers are flooded with data Tremendous quantity and multiple scales of data Difficult to collect, store, and manage How can we distill meaningful knowledge from data?
3 Data is the 4th Paradigm Producing an avalanche of high resolution digital data All (or most) of the data needs to be accessible over a long period of time Much of the data is not reproducible Example NEES project Structure or sample destroyed through testing Very expensive to rebuild for more tests
4 Data, data every where We are surrounded by data that we want, but it is difficult to find the information that we need Water, water every where, Nor any dro to drink. Samuel Taylor Coleridge, Rime of the Ancient Mariner Private, shared, and public data repositories Files on your computer Group documents and files Experimental results Published papers Data are scattered across many systems and devices Personal computer, old diskettes in a box, several systems, Old computer systems The Rime of the Ancient Mariner: Plate 32: The Pliot, by Gustave Doré
5 Need for Data Education Data is the 4th paradigm of Science and Engineering We are losing valuable data every day The techniques we were taught to maintain a lab notebook has not been effectively transferred to computer based data collection and registration systems. So much data is available and collected today, it is not possible to keep it on paper anymore.
6 Two Examples of Data Intensive Science Two large-scale science and engineering projects illustrate the problems related to data intensive science National Science Foundation George E. Brown Network for Earthquake Engineering Simulation (NEES) Purdue operates the headquarters for the NEEScomm, the community of NEES research facilities The Compact Muon Solenoid project Purdue operates a Tier-2 CMS center
7 NSF Network for Earthquake Engineering Simulation (NEES) Safer buildings and civil infrastructure are needed to reduce damage and loss from earthquakes and tsunamis To facilitate research to improve seismic design of buildings and civil infrastructure, the National Science Foundation established NEES NEES Objectives Develop a national, multi-user, research infrastructure to support research and innovation in earthquake and tsunami loss reduction Create an educated workforce in hazard mitigation Conduct broader outreach and lifelong learning activities
8 Vision for NEES Facilitate access to the world's best integrated network of state-of-the art physical simulation facilities Build a cyber-enabled community that shares ideas, data, and computational tools and models. Promote education and training for the next generation of researchers and practitioners. Cultivate partnerships with other organizations to disseminate research results, leverage cyberinfrastructure, and reduce risk by transferring results into practice.
9 NEES Research Facilities NEES has a broad set of experimental facilities Each type of equipment produces unique data Located at 14 sites across the United States Shake Table Tsunami Wave Basin Large-Scale Testing Facilities Centrifuge Field and Mobile Facilities Large-Displacement Facility Cyberinfrastructure
10 Oregon State University University of Minnesota University of Illinois- Urbana University of California Berkeley University of California Davis https://www.nees.org University of Buffalo University of California Santa Barbara Cornell University University of California Los Angeles Rensselaer Polytechnic Institute University of California San Diego University of Nevada Reno University of Texas Austin Lehigh University
11 Large-Scale Testing Facilties Lehigh University Reaction wall, strong floor dynamic actuators UC-Berkeley Reconfigurable Reaction Wall University of Illinois Urbana-Champaign Multi-Axial Full-Scale Sub-Structured Testing & Simulation (MUST-SIM) University of Minnesota Reaction walls Multi-Axial Subassemblage Testing (MAST) Images: Univ of Minnesota
12 NEEShub at Nees.org
13 Compact Muon Solenoid Project Another example of a big data project Two primary computational goals Move detector data from Large Hadron Collider at CERN to remote sites for processing Examine detector data for evidence of Higgs boson ~15 PB/yr data Applications used by CMS are not inherently parallel Data is split up and distributed across nodes Embarrassingly parallel
14 CMS Project Overview CERN Large Hadron Collider Project (LHC) Particle accelerator and collider largest in the world 17 mile circumference tunnel Providing evidence to support the existence of the Higgs boson Six detector experiments at the LHC Atlas, CMS, LHCb, ALICE, TOTEM, LHCf Compact Muon Solenoid (CMS) Very large solenoid with 4 Tesla magnetic field Earth s magnetic field 60 x 10^-6 Tesla
15 CMS Detector
17 Purdue CMS Tier-2 Center Computing Infrastructure ~10,000 computing cores within the Purdue University Community Cluster program Purdue recently (June 18) announced the Conte Supercomputer Fastest university-owned supercomputer in the United States 3 PB of disk storage running Hadoop Sharing a 100 Gb/sec network uplink to Indianapolis and Chicago Ultimately connecting to Fermi National Lab in Chicago Provided 14% of all Tier-2 computing globally in 2012
18 Purdue CMS Tier-2 Center Physicists from around the world submit computational jobs to Purdue Data is copied from the Tier-1 to Purdue storage on user request Simulation codes also run at Purdue, with results pushed up to Tier-1 center or other Tier-2s. International data sharing Data interoperability is designed into the project from the beginning. There is one instrument (the CMS detector), which greatly simplified the sharing and reuse of data compared with a project like NEES
19 Challenges involved in Big Data Performance at Scale How can we effectively match data performance with HPC capabilities? How can we ensure good reliability of these systems? Data Curation Challenges What should we preserve, how should we preserve it, and how can we ensure the long-term viability of the data? Disciplinary Sociology and Cyberinfrastructure How can we effectively promote and support the adoption and use of new technologies? How can we foster the development of new disciplinary practices focused on the long-term accessibility of data?
20 Performance at Scale Petaflop scale systems are now available for use by researchers Example: Purdue Conte system announced this week (Rmax 943 TF, Rpeak Petaflops) Conte was built with 580 HP ProLiant SL250 Generation 8 (Gen8) servers, each incorporating two Intel Xeon processors and two Intel Xeon Phi coprocessors, integrated with Mellanox 56Gb/S FDR InfiniBand. Conte has 580 servers (570 at the time of testing) with 9,120 standard cores and 68,400 Phi cores, for a total of 77,520 cores. Big data analytics coupled with petascale systems requires high bandwidth storage systems Avoid wasteful and expensive CPU stalls Scaling up is along two axes: Large volume of data (example: CMS Project) Large variety and number of files (example: NEES project)
21 Curation Challenges Data production rate is tremendous Volume of data is growing over time Sensor sampling rate increasing High definition video Managing data transfer Time required to upload and download data is growing Upload and download time can take a lot of time if there are network bottlenecks Ensuring data integrity Filtering, cleaning, and calibration is often needed before upload and curating data The community needs to also retain the raw data in case an error is made or in case a researcher can later distill further insights from the data.
22 Curation Challenges File type management Data is stored in files through the intermediary of an application This means that the information in the data will be encoded into some kind of format It s difficult (if not impossible) to restrict the file formats used by the research community As these applications change (or disappear) over time, the information encoded in the data may become stranded Risk of stranded data When the file format cannot be precisely identified, then we don t know which application can be used as an intermediary for reading the information encoded in the data. This leads to stranded data that is useless.
23 Curation Challenges Linking computation with data and archived data Will need the ability to quickly search archived data much more detailed that what Google can deliver How can we quickly discover, convert, and transfer archived data to be close to the user and to computation? (especially HPC) Need to match data I/O capabilities with growth in the number of CPU cores and core speed.
24 Long-term accessibility We have data in the NEEShub from the 1970s Science: Rescue of Old Data Offers Lesson for Particle Physicists by Andrew Curry (Feb 2011) Described the need to find old, almost lost data for a physics experiment from the 1980s The data will need to remain viable and accessible for years into the future
25 Discipline Sociology Sociological factors in data curation Disciplinary differences in how data are archived, how to value archived data, and determining what is worth retaining Who determines what is worth keeping? What is the practice in the specific discipline? International standards and practices in metadata tagging, representing numbers, and even character sets NEES is working with partners in Japan and China we need to determine how to represent their data in a common standard framework Terminology for numbers (, vs.., lakh vs. 100,000) Changing the behavior of scientists to value curation and long-term accessibility
26 Managing Curation at Scale How can we efficiently use data curator s time? NEES now has 1.8M files, what will happen in 3 more years? How can we manage 10M files with a limited curation staff? For NEES,we are using the OAIS model as a guideline for designing a curation pipeline for curating NEES data The OAIS model is proving to be a useful model for thinking about how to undertake data curation We are developing a curation pipeline to help automate curation for the many files in the NEES Project Warehouse
27 Data Analytics There are technologies available today that can be used to provide solutions to these problems High performance computing Parallel file systems Map Reduce/Hadoop A sustainable solution requires more than a set of technologies An effective data cyberinfrastructure involves both sociological and technological components. What is needed to educate and train researchers to effectively learn to use new technologies?
28 Our approach Developing a joint research and education and program in big data analytics between the University of Stavanger and Purdue University and AMD Research. Chunming Rong, Tomasz Wlodarczyk (Stavanger) Thomas Hacker, Ray Hansen, Natasha Nikolaidis (Purdue) Greg Rodgers (AMD Research) Funded by SIU: Strategic Collaboration on Advanced Data Analysis and Communication between Purdue University and University of Stavanger Developing a semester long joint course in HPC and Big Data Analytics, and a short summer course (to be delivered next week)
29 Planned Course Objectives Students will learn to put modern tools to use in order to do data analysis of large and complex data sets. Students will be able to: design, construct, test, and benchmark a small data processing cluster (based on Hadoop) Demonstrate knowledge of MapReduce functionalities through the development of a MapReduce program Understand Hadoop job tracker, task tracker, scheduling issues, communications, and resource management. Construct programs based on MapReduce paradigm for typical algorithmic problems Use functional programming concept to describe data dependencies and analyze complexity of MapReduce programs
30 Planned Course Objectives Algorithms Understand algorithmic complexity of the worst case, expected case, and best case running time (big-oh notation), and the orders of complexity (e.g. N, N^2, Log N, NP-Hard) Examine a basic algorithm and identify the algorithmic complexity order File Systems Describe the concepts of a distributed file system, how it differs from a local file system, the performance of distributed file systems. Describe a parallel file system, the performance advantages possible through the use of a parallel file system, and the inherent reliability and fault tolerance mechanisms needed for parallel file systems. Examples include OrangeFS and Lustre understand peak and sustained bandwidth rates understand the differences between RDBMS, data warehouse, unstructured big data, and keyed files.
31 Short Course Format Lecture in the morning followed by lab in the afternoon Labs are built on a set of Desktop PCs running Hadoop in an RHEL6 virtual machine running on top of VMware Using pfsense (open source firewall) to create a secure network connection from the instruction site to the computers running Hadoop Working to refine the network and lab equipment setup based on our experiences delivering the short course next week.
32 Short Course Day 1 Topics Lecture Introduction and motivation for the course History of HPC, big data, Moore's Law. Science domain areas, and problems in each of those areas that motivate the need for this. Where are we today, and what is the projected need later? How are things driven by increases in computing power? Definition of big data, big compute, why we need both combined Mixture of trends, principles, and implementation in historic context that students should understand. Parallel application types Introduction to MapReduce Dataflow within MapReduce with plug-in Labs The hadoop command, HDFS, and Linux basics Hadoop basic examples from lectures
33 Short Course Day 2 Topics Lectures Introduction to MapReduce, continued Combiners More complex MapReduce example (search assist) Hadoop Architecture Motivation for Hadoop Hadoop building blocks (name node, data node, etc.) Fault tolerance and failures, replication, and data aware scheduling. Main components (HDFS, MapReduce, modes (local, distributed, pseudo distributed), etc.) HDFS GUI Labs We will use combiners and multiple reducers to improve performance. We will look at network traffic and data counters to evaluate. Students will evaluate the performance improvement for each optimization of MapReduce program. The advanced student will gather network and data statistics to explain why each phase got better.
34 Short Course Day 3 Topics Lectures Hadoop Architecture, continued Comparison of HDFS with other Parallel File System architectures (GoogleFS, Lustre, OrangeFS), and how Hadoop differs from these systems Chaining MapReduce jobs Mapreduce Algorithms: K-means or other algorithms Schemas for unstructured data using Hive Introduction to data organization. Why are we concerned about data organization? What are the impacts of poor organization on performance and correctness? Data organization: Level of data organization - data structure, file level, cluster level, data parallelization, organization level. How do we deal with large sequential files from a performance perspective and how it would be represented in parallel file system (e.g. HDFS) Lab Hive
35 Expected Outcomes Provide education and training to researchers to allow them to effectively think about big data to effectively use the technologies in their research and daily work. Improved data collection and management practices by researchers Development of new techniques for collaboration on a joint course across the Atlantic with a shared lab infrastructure for lab assignments.
36 Conclusions There is a need for data intensive training and education for scientists and engineers Effectively use existing technologies Develop new disciplinary practices for annotating and preserving valuable data Understand the critical need for data curation for the viability and long-term accessibility of data We are developing a education and research program focused on these issues Short course Semester length joint course at University of Stavanger and Purdue University Holding a symposium at the CloudCom conference in December DataCom - Symposium on High Performance and Data Intensive Computing Thomas Hacker, Purdue Univ., USA Tomasz Wiktor Wlodarczyk, University of Stavanger, Norway DataCom is organized under CloudCom as two tracks Big Data HPC on Cloud