exascale(compu.ng(and( materials(discovery( ·...
TRANSCRIPT
Exascale Compu.ng and Materials Discovery
Bruce Harmon
Ames Laboratory, USDOE
and Iowa State University
LAUSANNE -‐ May 25, 2011
Outline: 1. Exascale (when? and how?)
2. Tipping point for Materials
Discovery?
3. Local (Ames Lab) Example
Performance for top 10 Supercomputers over Gme.
Year From Report: ExaScale Compu.ng Study: Technology Challenges in Achieving Exascale Systems, 2008. DARPA
DOE -‐ Exascale Ini.a.ve High Level Targets
• The Exascale Ini.a.ve targets plaSorm deliveries in 2018 and a robust Exascale simula.on environment for the science exemplars by 2020
• Co-‐development of hardware, system soTware, programming model and applica.ons require intermediate (100-‐200 PF/s) plaSorms in 2015
DOE Exascale Ini.a.ve Technical Roadmap
DOE Exascale Ini.a.ve Technical Roadmap
Poten.al System Parameters for Exascale
Systems 2009 2011 2015 2018
System peak 2 Peta 20 Peta 100-‐200 Peta 1 Exa
System memory 0.3 PB 1.6 PB 5 PB 10 PB
Node performance 125 GF 200GF 200-‐400 GF 1-‐10TF
Node memory BW 25 GB/s 40 GB/s 100 GB/s 200-‐400 GB/s
Node concurrency 12 32 O(100) O(1000)
Interconnect BW 1.5 GB/s 22 GB/s 25 GB/s 50 GB/s
System size (nodes) 18,700 100,000 500,000 O(million)
Total concurrency 225,000 3,200,000 O(50,000,000) O(billion)
Storage 15 PB 30 PB 150 PB 300 PB
IO 0.2 TB/s 2 TB/s 10 TB/s 20 TB/s
MTTI days days days O(1 day)
Power 6 MW ~10MW ~10 MW ~20 MW
Power Consump.on • Barriers – Power is leading design constraint for compu.ng
technology – Target ~20MW, es.mated > 100MW required for
Exascale systems (DARPA, DOE) – Efficiency is industry-‐wide problem (IT
technology >2% of US energy consump.on and growing)
• Technical Focus Areas – Energy efficient hardware building blocks (CPU,
memory, interconnect) – Novel cooling and packaging – Si-‐Photonic Communica.on – Power Aware Run.me SoTware and Algorithms
• Technical Gap – Need 5X improvement in power efficiency over
projec.ons that include technological advancements
Possible Leadership class power requirements
From Peter Kogge (on behalf of Exascale Working Group), “Architectural Challenges at the Exascale Fron.er”, June 20, 2008
Slide 6 DOE Exascale Ini.a.ve Technical Roadmap
Desired
Projected including industry BAU improvements
Reliability and Resilience • Barriers – Number of system components increasing faster
than component reliability – Mean Gme between failures of minutes or seconds
for exascale – Silent error rates increasing – No job progress due to fault recovery if we use
exis.ng checkpoint/restart
• Technical Focus Areas – Improved hardware and soTware reliability
• Beler RAS collec.on and analysis (root cause) • Greater integra.on
– Fault resilient algorithms and applica.ons – Local recovery and migra.on
• Technical Gap – Need 1000X improvement in MTTI so that
applica.ons can run for many hours. Goal is 10X improvement in hardware reliability. Local recovery may and migra.on may yield another 10X. However, for exascale, applica.ons will need to be fault resilient.
Slide 7 DOE Exascale Ini.a.ve Technical Roadmap
The image cannot be displayed. Your computer may not have enough memory to open the image, or the image may have been corrupted. Restart your computer, and then open the file again. If the red x still appears, you may have to delete the image and then insert it again.
.me
2000 2015 2018
Time to checkpoint grows larger as problem size increases
MTTI grows smaller as scale increases
By exascale checkpoint/restart no longer viable
Effec.ve applica.on u.liza.on (including checkpoint overhead) at 3 rates of hardware failure
Power
Memory and Storage Bandwidth
Reliability and Resilience
System SoXware Scalability
Programming Models and Environments
Slide 8 DOE Exascale Ini.a.ve Technical Roadmap
2012 2013 2014 2015 2017 2018 2019 2020
10 Peta 1 Exa 100 Peta
2016
Technology Roadmap
Exascale Science
Memory BW 10x
Demonstrate > 3X power efficiency gain over 2012
SW scalability to 10M threads
3D chip-‐level integraGon
Improved hardware and soXware reliability
Local fault recovery
New IO technology
Latency tolerant algorithms
New programming model
Improved interconnect technology
Demonstrate 10X power efficiency gain over 2015
Fault Tolerant ApplicaGons
SW scalability to 10M threads
ApplicaGon scalability to 100M threads
ApplicaGon scalability to 100M threads
Benefits from Investments Applica.ons • The solu.on to the na.on’s 5-‐6 top science and security challenges is the ul.mate benefit of the exascale ini.a.ve. Examples include na.onal and energy security, climate change, and advanced materials.
Programming Models and Environments • New programming environments will enable scien.sts across DOE to use these systems (not just a
handful of hero HPC programmers). • Scalable, resilient system soTware will enable exascale systems to be effec.ve tools for science and
improve resilience of smaller systems across the porSolio. • New tools will shorten the development cycle for tackling new na.onal challenges,
Computer Systems • The benefits of technology advances in resilience and power efficiency required for an exascale
system will also trickle down to smaller systems resul.ng in reduced na.onal energy demand and broad impact to science in the U.S.
• The strategy of staged plaSorms provides a path for applica.on development and the resources for intermediate science breakthroughs
Slide 9 DOE Exascale Ini.a.ve Technical Roadmap
Mo.va.on / Driving Forces
MATERIALS DISCOVERY Data mining and accelerated electronic structure theory
as a tool in the search for new func.onal materials C. Or.z 1, O. Eriksson, M. Klintenberg Uppsala University, Uppsala, Sweden Computa.onal Material Science 44 (2009) 1042-‐1049
Evolu.onary crystal structure predic.on as a tool in materials design Artem R Oganov and ColinW Glass J. Phys.: Condens. Maler 20 (2008) 064210
Materials Informa.cs: More Than Collec.ng Data Kim F. Ferris, Pacific Northwest Na.onal Laboratory Dumont M. Jones, Proximate Technologies, LLC. Talk.
Overview • What is materials informaGcs
– System architecture – General definiGon – Compare and contrast with bioinformaGcs
• Basic Approach to Materials InformaGcs – Flow chart
• Database development • Testbed environment • StaGsGcal development • ProjecGon and new candidate generaGon
• Benefits of InformaGcs vs. Linear Searches • Examples • Goals of Structured Materials Discovery
• Provide a consistent basis for electronic and structural informaGon derived from first-‐principles computaGon with exisGng empirical and computaGonal data.
• Generate materials informaGcs toolkit for the development of design rules based on staGsGcal learning theory, simulaGon, and empirical evidence
• Develop a materials database resource which can be used as a basis for future materials development or verificaGon (potenGally within and outside PNNL)
• Develop a signature laboratory capability based upon informaGon theory techniques for new materials development.
Summary of Expected Outcomes
Current Ames Lab Example
Mo.vated by need Cri.cal Materials Concern Desire to find NEW high performance magne.c materials containing no rare earth elements Use first principles calcula.ons and perform massive searches on supercomputers Strong interac.on with experimental groups
GeneGc algorithm (GA) for global structure opGmizaGons
• Global structure opGmizaGon is a big challenge
• TradiGonal approaches use simulated annealing
• GeneGc algorithm approach based on a physical representaGon (i.e., atomic coordinates) is simple but performs befer than simulated annealing
• GA approach has been applied to – Atomic clusters – Surfaces and interfaces – Nanowires
ma.ng opera.on-‐clusters
• Pick two clusters from the pool.
• Rotate them randomly • Cut through the z=0 plane • Paste atoms above z=0
from A and those below z=0 from B
• ShiT cluster along z-‐direc.on to have correct match in number of atoms
D. Deaven and K.M Ho Molecular Geometry Op6miza6on with a Gene6c Algorithm PRL 75 288 (1995)
Use of GA for crystal structure predic.on
• A.R. Oganov and C. W. Glass, “Crystal structure predic.on using ab ini6o evolu.onary techniques: Principles and applica.ons”, J. Chem. Phys. 124, 244704(2006).
• G. Trimarchi and A. Zunger, “Finding the lowest-‐energy crystal structure star.ng from randomly selected lavce vectors and atomic posi.ons: first-‐principles evolu.onary study of the Au–Pd, Cd–Pt, Al–Sc, Cu–Pd, Pd–Ti, and Ir–N binary systems”, J. Phys.: Condens. Ma@er 20, 295212 (2008).
• G. Trimarchi, A. J. Freeman and A. Zunger, “Predic.ng Stable Stoichiometries of Compounds via Evolu.onary Global Space-‐group Op.miza.on”, Phys. Rev. B, 80, 092101 (2009).
• A.R. Oganov, J.H. Chen ,C. Gav, Y.Z. Ma, Y. M. Ma, C.W. Glass, Z.X. Liu , T. Yu, O.O. Kurakevych and V.L. Solozhenko, “Ionic high-‐pressure form of elemental boron”, Nature, 457,863(2009).
Some relevant references
Bolleneck of computa.onal crystal structure predic.on
• M. Ji, C. Z. Wang, and K. M. Ho, “Comparing efficiencies of gene.c and minima hopping algorithms for crystal structure
predic.on”, Phys. Chem. Chem. Phys, 12, 11617-‐23 (2010)
• Searches using empirical poten.als are fast but suffer from inaccuracies which can lead the search to wrong structures.
• GA searches using ab-‐ini.o calcula.ons are restricted because the computa.onal effort needed to effec.vely sample the structure configura.on space is extremely demanding for any but the smallest unit cells.
• A breakthrough is needed which can address both the speed and effec.veness of the search with an accurate descrip.on of interatomic interac.ons in the system.
Adap.ve GA • Ji Min, C. Z., K. M. • Most of the computer .me is spent in simula.ng the
relaxa.on of false structural candidates. • Employ auxillary poten.als to give an es.mate of the
approximate energy ordering of the different compe.ng geometries. Star.ng with some educated guesses, the accuracy of the auxillary poten.als can be improved by adap.ve adjustments during the course of the GA process
• Method combines the speed of empirical poten.als with the accuracy of ab-‐ini.o calcula.ons
Adap.ve GA atomic structure search of FeCo alloy
Manh Cuong Nguyen, Min Ji, Cai-‐Zhuang Wang and Cai-‐Ming Ho
Relaxa.on in DFT
• Lowest energy structures with energy window of 5 meV/atom.
• Structures are fully relaxed.
Relaxa.on in DFT
• Lowest energy structures with energy window of 5 meV/atom.
• Structures are fully relaxed.
Relaxa.on in DFT
• Lowest energy structures with energy window of 5 meV/atom.
• Structures are fully relaxed.
• GA atomic structure search “from scratch” was implemented for FeCo and FeCoW systems.
• Low energy structures of FeCo and FeCoW systems have a BCC underlying palern, consistent with experiment.
• Ground states of FeCo systems are highly degenerate. • Low energy structures of FeCoW system have 2 main palerns of Co atoms: triangle and line.
FIND:
Conclusions
Petascale compu.ng will greatly enhance theore.cal capability to search for new cri.cal material candidates. Strong partnering with synthesis and charateriza.on groups, promises to greatly accelerate the discovery of new materials.