1 2014 fujitsu december 15 th 2014 high performance computing

45
1 2014 Fujitsu December 15 th 2014 High Performance Computing

Upload: rachel-douglas

Post on 21-Dec-2015

216 views

Category:

Documents


1 download

TRANSCRIPT

  • Slide 1
  • 1 2014 Fujitsu December 15 th 2014 High Performance Computing
  • Slide 2
  • 2 2014 Fujitsu Agenda Introduction Why Parallel Why not Parallel Speedup How the Code Looks Like Fujitsu Value Proposition to KAU Benefits to Society
  • Slide 3
  • 3 2014 Fujitsu Introduction In the field of research, High Performance Computing (HPC) is the use of hardware, software, tools and programming techniques to accelerate research computation, which in turn will enable the execution of large cutting-edge research simulation that accelerates new discoveries.
  • Slide 4
  • 4 2014 Fujitsu Sequential Processing Matrix Addition Example
  • Slide 5
  • 5 2014 Fujitsu Sequential Processing Matrix Multiplication Example
  • Slide 6
  • 6 2014 Fujitsu Sequential Processing What will happen if the matrix size is 1000,000,000 x 1000,000,000 ? If each addition operation needs 1 Microsecond then we need more than 3000 Years to finish the computation
  • Slide 7
  • 7 2014 Fujitsu Sequential Processing Standalone Computers are not able to face Big Data Analysis and Processing
  • Slide 8
  • 8 2014 Fujitsu Parallel Processing Matrix Addition Example
  • Slide 9
  • 9 2014 Fujitsu Parallel Processing Matrix Multiplication Example
  • Slide 10
  • 10 2014 Fujitsu Agenda Introduction Why Parallel Why not Parallel Speedup How the Code Looks Like Fujitsu Value Proposition to KAU Benefits to Society
  • Slide 11
  • 11 2014 Fujitsu Why Parallel It is not always obvious that a parallel algorithm has benefits, unless we want to do things faster: doing the same amount of work in less time bigger: doing more work in the same amount of time Both of these reasons can be argued to produce better results, which is the only meaningful outcome of program parallelization
  • Slide 12
  • 12 2014 Fujitsu Faster, Bigger! There is an ever increasing demand for computational power to improve the speed or accuracy of solutions to real-world problems through faster computations and/or bigger simulations Computations must be completed in acceptable time (real-time computation), hence must be fast enough
  • Slide 13
  • 13 2014 Fujitsu Faster, Bigger! An illustrative example: a weather prediction simulation should not take more time than the real event Suppose the atmosphere of the earth is divided into 510 8 cubes, each 111 mile and stacked 10 miles high It takes 200 floating point operations per cube to complete one time step 10 4 time steps are needed for a 7 day forecast (time step = 1 min) Then 10 15 floating point operations must be performed (510 8 x 200 x 10 4 ) This takes 10 6 seconds (= 10 days) on a 1 GFLOP machine
  • Slide 14
  • 14 2014 Fujitsu Grand Challenge Problems Big problems A Grand Challenge problem is a problem that cannot be solved in a reasonable amount of time with todays computers Examples of Grand Challenge problems: Applied Fluid Dynamics Meso- to Macro-Scale Environmental Modeling Ecosystem Simulations Biomedical Imaging and Biomechanics Molecular Biology Molecular Design and Process Optimization Fundamental Computational Sciences Nuclear power and weapons simulations
  • Slide 15
  • 15 2014 Fujitsu Physical Limits Which tasks are fundamentally too big to compute with one CPU? Suppose we have to calculate in one second for (i = 0; i < ONE_TRILLION; i++) z[i] = x[i] + y[i]; Then we have to perform 3x10 12 memory moves per second If data travels at the speed of light (3x10 8 m/s) between the CPU and memory and r is the average distance between the CPU and memory, then r must satisfy 310 12 r = 310 8 m/s 1 s which gives r = 10 -4 meters To fit the data into a square so that the average distance from the CPU in the middle is r, then the length of each memory cell will be: 210 -4 m / (310 6 ) = 10 -10 m which is the size of a relatively small atom!
  • Slide 16
  • 16 2014 Fujitsu Agenda Introduction Why Parallel Why not Parallel Speedup How the Code Looks Like Fujitsu Value Proposition to KAU Benefits to Society
  • Slide 17
  • 17 2014 Fujitsu Why not Parallel Bad parallel programs can be worse than their sequential counterparts Slower: because of communication overhead Scalability: some parallel algorithms are only faster when the problem size is very large Understand the problem and use common sense! Not all problems are amenable to parallelism Some algorithms are inherently sequential for (i=1; i< 1000000; i++) X[i]=X[i-1] + Y[i]; We can find that there is a dependency that prevents parallelism
  • Slide 18
  • 18 2014 Fujitsu Agenda Introduction Why Parallel Why not Parallel Speedup How the Code Looks Like Fujitsu Value Proposition to KAU Benefits to Society
  • Slide 19
  • 19 2014 Fujitsu Speedup Definition: the speedup of an algorithm using P processors is defined as S P = T s / T P Where: T s is the execution time of the best available sequential algorithm and T P is the execution time of the parallel algorithm
  • Slide 20
  • 20 2014 Fujitsu Amdahls Law s (1- ) T s TsTs P S P = T s / T P + (1- ) P 1 S P = ( s ) + (1- ) T s P TsTs S P =
  • Slide 21
  • 21 2014 Fujitsu Parallel Search
  • Slide 22
  • 22 2014 Fujitsu Agenda Introduction Why Parallel Why not Parallel Speedup How the Code Looks Like Fujitsu Value Proposition to KAU Benefits to Society
  • Slide 23
  • 23 2014 Fujitsu How the Code Looks Like Sequential Matrix Addition for i = 1 to M for j = 1 to N C[i,j] = A[i,j] + B[i,j] Parallel Matrix Addition for i = 1 to M for j = 1 to N C[i,j] = A[i,j] + B[i,j] #pragma omp for Matrix Addition Example
  • Slide 24
  • 24 2014 Fujitsu Agenda Introduction Why Parallel Why not Parallel Speedup How the Code Looks Like Fujitsu Value Proposition to KAU Benefits to Society
  • Slide 25
  • 25 2014 Fujitsu HPC at KAU Many scientific fields have adopted in the provisioning and support of research computation. In these environments, departments and research groups are providing local resources to support their researchers, whereas the IT team is focusing on providing central HPC resources and services to support all staff and student researchers across campus. IT Team Scientists and Research Students Focus on Solving Scientific Problems Focus on Providing Sustainable Central HPC Service
  • Slide 26
  • 26 2014 Fujitsu HPC at KAU Fujitsu is currently working with KAU in order to provide researchers with impeccable HPC resources to fulfill their increasing need and meet or even go beyond their research target. In that sense Fujitsu will collaborate closely and strongly with KAU in order to ensure that : The use of HPC is propagated to the whole research communities within KAU KAU researchers become experienced HPC users KAU IT team become knowledgeable in HPC technologies and management
  • Slide 27
  • Fujitsu value proposition to KAU 27 A proven set of services ( support and professional) delivered by Fujitsu HPC experts The Innovative approach of HPC Campus aligned to the strategy of KAU in promoting HPC in the University and the Kingdom The involvement of the whole Fujitsu Corporation with a full commitment and the access to all HPC worldwide expertise A proven set of services ( support and professional) delivered by Fujitsu HPC experts The Innovative approach of HPC Campus aligned to the strategy of KAU in promoting HPC in the University and the Kingdom The involvement of the whole Fujitsu Corporation with a full commitment and the access to all HPC worldwide expertise Fujitsu solution Implement a solid foundation to support R&D activities and allow scientists to focus on their research challenges Sound and dedicated service model End users access and ease of use Enlarge HPC user community Generate ROI visible to the whole Kingdom Sound and dedicated service model End users access and ease of use Enlarge HPC user community Generate ROI visible to the whole Kingdom KAU requests
  • Slide 28
  • Solution Building blocks Hardware (servers, network, storage ) OS (RedHat EL, CentOS) HPC middleware (admin, workload mgr, MPI, Libs ) User interface (Web portal) User Applications Infrastructure layer (machine room, cooling, electricity )
  • Slide 29
  • Commercial in confidence - Copyright 2012 FUJITSU LIMITED Services Overview KAU requirementsFujitsus Service delivery A fully supported and maintained premier HPC systemDedicated Staff to fully manage and operate the facility Ability to operate and manage the HPC systems with internal teams Training and Knowledge Transfer Program for (End Users, Application Developers, System Administrators, and Operators) Porting/validation of existing applications on the new service Migration of the existing implementations to the new HPC facility Optimizing/Extending existing applicationsSupport in Code optimization Ability to incorporate remotely located compute islands on campus Migration and integration Services Positioning KAU as an HPC competence centre in the Kingdom and the GCC Establishing a collaboration scheme between KAU and International research organization. HPC global awareness program Increase institutional collaboration worldwideWell structured collaboration activities management to Increase the scope of research project with international organizations Promote the use of HPC across other areas of science and research Comprehensive Outreach and marketing Program Link remote organisations resources to KAUs HPC service Campus cloud Implementation
  • Slide 30
  • Commercial in confidence - Copyright 2012 FUJITSU LIMITED Service organization The whole operations are managed by on-site team The skill-set (on site or mobile) needed to deliver the services are covering all the areas of expertise needed to deliver the services : HW engineer (on site and mobile) for HW maintenance SW engineer (on site and mobile) for SW maintenance Application specialist (on site and mobile) for application support Technical experts used for 3 rd level support (mobile), either Fujitsu or 3rd party experts HPC consultants used in delivering professional services 30 On-site operation and support Project Manager HW engineer SW engineer Application expert HW engineer SW engineer Application expert Intel support Cisco support NetApp support Altair support Fujitsu product support HPC SW Technical group HPC HW technical support Mobile HW engineer 3rd party experts Fujitsu experts
  • Slide 31
  • Single service interface with worldwide backing 31 Fujitsu HPC Tokyo Toulouse London Munich KAU Service Desk End-to-End services are delivered using the Worldwide expertise of Fujitsu in HPC. This covers the complete scope of HPC components ( HW, SW, Applications) End-to-End services are delivered using the Worldwide expertise of Fujitsu in HPC. This covers the complete scope of HPC components ( HW, SW, Applications) San Jose Fujitsu service desk Single point of contact Fujitsu service desk Single point of contact
  • Slide 32
  • Commercial in confidence - Copyright 2012 FUJITSU LIMITED Agenda Introduction Why Parallel Why not Parallel Speedup How the Code Looks Like Fujitsu Value Proposition to KAU Benefits to Society
  • Slide 33
  • Commercial in confidence - Copyright 2012 FUJITSU LIMITED Samples of the Benefits to Society There are a whole range of benefits that could be gained by society through the application of HPC: HPC could considerably increase oil recovery with more accurate seismic modeling of oil reservoirs. Currently, the uncertainties in the seismic models can lead to errors in drilling that both decrease output and increase environmental impact. HPC is being used to design efficient wind and wave turbines, helping to harness renewable energy sources. HPC could be used to model the spread of epidemics, enabling public health officials to intervene appropriately to halt the expansion of life-threatening diseases. HPC could lead to a revolution in medical procedures and devices as well as product safety for a variety of consumer products by simulating virtual humans of all shapes and ages. 33
  • Slide 34
  • Commercial in confidence - Copyright 2012 FUJITSU LIMITED Other Applications of HPC Simulation and Modeling
  • Slide 35
  • Commercial in confidence - Copyright 2010 FUJITSU LIMITED Other Applications of HPC Bioinformatics
  • Slide 36
  • Commercial in confidence - Copyright 2010 FUJITSU LIMITED Oil and Gas AerodynamicsFluid Mechanics Soil Mechanics GeophysicsRemote Sensing Climate Research Oceanography Aerospace Other Applications of HPC Big Data Analysis
  • Slide 37
  • Commercial in confidence - Copyright 2012 FUJITSU LIMITED Agenda Introduction Why Parallel Why not Parallel Speedup How the Code Looks Like Fujitsu Value Proposition to KAU Benefits to Society
  • Slide 38
  • Commercial in confidence - Copyright 2012 FUJITSU LIMITED Deliverables HPC facility - fully operated and managed by KAU members. Professional HPC CoE High quality research work (research projects, scientific papers, etc.) Sustainable HPC facility with an ambitious research plan in the field of HPC 38 HPC Center of Excellence Publications ISI SCOPUS ELSEVIER Springer HPC Research Roadmap Experts Research Projects Funds
  • Slide 39
  • Commercial in confidence - Copyright 2012 FUJITSU LIMITED 39 380 Nodes 9120 Cores 112 Nodes 2688 Cores 2 NVIDIA 48 Core 2 Nodes 48 Core 96 GB 256 GB 96 GB
  • Slide 40
  • Commercial in confidence - Copyright 2012 FUJITSU LIMITED 40
  • Slide 41
  • Commercial in confidence - Copyright 2012 FUJITSU LIMITED How can we cooperate? Deploying the already existing software systems on the new platform. Implementing algorithms from scratch. Optimizing the existing codes. Working on much more larger data sets. Conducting professional training. Conducting professional training Initial Training Plan 41
  • Slide 42
  • Commercial in confidence - Copyright 2012 FUJITSU LIMITED Thank You 42
  • Slide 43
  • Commercial in confidence - Copyright 2010 FUJITSU LIMITED We have a dream Having a new parallel computing paradigm that can be used by the researchers and the scientific community in the field of HPC
  • Slide 44
  • Commercial in confidence - Copyright 2010 FUJITSU LIMITED 44
  • Slide 45
  • Commercial in confidence - Copyright 2010 FUJITSU LIMITED How the Code Looks Like Sequential Matrix Addition for i = 1 to M for i = 1 to N C[i,j] = A[i,j] + B[i,j] Parallel Matrix Addition for i = 1 to M for i = 1 to N C[i,j] = A[i,j] + B[i,j] #pragma omp for Matrix Addition Example