From Research to Large-Scale HPC Applications

Size: px

Start display at page:

Download "From Research to Large-Scale HPC Applications"

Ethelbert Norris
8 years ago
Views:

1 From Research to Large-Scale HPC Applications The Paradigm of Geophysical Imaging and Prospecting Ernesto Bonomi Energy & Environment Program CRS4, Pula, Italy

2 Seismic Imaging: a remote subsurface echography Subsurface reflec,vity map

3 Seismic Imaging Wishful Thinking Flop/s Improved physics vs algorithm complexity EFlop/s 100 ViscoelasBc FWI Petro- elasbc inversion 10 ViscoelasBc modeling Isotropic elasbc FWI 1 1 PFlop/s Isotropic/anisotropic elasbc modeling Isotropic elasbc RTM Isotropic acousbc FWI Isotropic/anisotropic acousbc modeling Isotropic/anisotropic acousbc RTM Isotropic/anisotropic one- way wave acousbc imaging Isotropic ray tracing Kirchhoff imaging A very fast increasing complexity with high costs in terms of: HW, maintenance, personnel and machine room Unsustainable energy consumption and heat generation which prevents frequency scaling

01 Isotropic/anisotropic acousbc modeling Isotropic/anisotropic acousbc RTM Isotropic/anisotropic one- way wave acousbc imaging Isotropic ray tracing Kirchhoff imaging

4 Seismic Imaging Wishful Thinking WaJ Flop/s 200M EFlop/s 20M 100 2M K 1 1 PFlop/s 20K 0.1 2K 0.01 Isotropic ray tracing Kirchhoff imaging Improved physics vs power algorithm consumpbon complexity ViscoelasBc FWI Petro- elasbc inversion ViscoelasBc modeling Isotropic elasbc FWI Isotropic/anisotropic elasbc modeling Isotropic elasbc RTM Isotropic acousbc FWI Isotropic/anisotropic acousbc modeling Isotropic/anisotropic acousbc RTM Isotropic/anisotropic one- way wave acousbc imaging HPC is power limited not area limited Cost of energy to run a cluster equals purchase cost in about 18 months Cost of provisioning a machine room with adequate power is typically many times cost of cluster What matters now is J/Flop or Flops/W Industrial objective: 200pJ/Flop in 45 nm (5GFlops/W) Efficient communication and memory circuits Efficient data and instruction supply Agile memory system

Isotropic/anisotropic elasbc modeling Isotropic elasbc RTM Isotropic acousbc FWI Isotropic/anisotropic acousbc modeling Isotropic/anisotropic acousbc RTM Isotropic/anisotropic one- way wave acousbc

5 HPC Trends in Seismics since the performance of a single CPU cannot increase like in the past, the dominant paradigm in computer architecture is in the form of Mul,core processors Impossibile visualizzare l'immagine. La memoria del computer potrebbe essere insufficiente per aprire l'immagine oppure l'immagine potrebbe essere danneggiata. Riavviare il computer e aprire di nuovo il file. Se viene visualizzata di nuovo la x rossa, potrebbe essere necessario eliminare l'immagine e inserirla di nuovo. General purpose graphics processors Dataflow Engines (FPGAs)

La memoria del computer potrebbe essere insufficiente per aprire l'immagine oppure l'immagine potrebbe essere danneggiata.

6 HPC Trends in Seismics since the performance of a single CPU cannot increase like in the past, the dominant paradigm in computer architecture is in the form of Mul,core processors Impossibile visualizzare l'immagine. La memoria del computer potrebbe essere insufficiente per aprire l'immagine oppure l'immagine potrebbe essere danneggiata. Riavviare il computer e aprire di nuovo il file. Se viene visualizzata di nuovo la x rossa, potrebbe essere necessario eliminare l'immagine e inserirla di nuovo. General purpose graphics processors Dataflow Engines (FPGAs)

7 HPC Trends in Seismics since the performance of a single CPU cannot increase like in the past, the dominant paradigm in computer architecture is in the form of Mul,core processors General purpose graphics processors Dataflow Engines (FPGAs)

in computer architecture is in the form of Mul,core

8 HPC Trends in Seismics In spite of their spectacular performance, multicores, GPUs and FPGAs obey to a restrictive programming paradigm, data parallelism or flow parallelism, that makes difficult the numerical solution of many algebraic problems

programming paradigm, data parallelism or flow parallelism,

9 HPC Trends in Seismics The In spite most of efficient their spectacular way to implement performance, PDEs multicores, on these frontier GPUs and architectures FPGAs obey is to to a restrictive approximate programming the solution paradigm, on a structured data grid parallelism no indirect or flow addressing parallelism, that makes difficult the numerical using solution an of explicit many time algebraic marching problems scheme no matrix inversion Drawbacks Only conditionally stable Limit on (me- step size Numerical dispersion problems Small (me steps or high- order schemes Boundary conditions often introduce irregularities that require additional coding effort (absorbing boundaries) The solver becomes riddled with conditionals making the parallel executions loosely synchronous and then inefficient

time algebraic marching problems scheme no matrix inversion Drawbacks Only conditionally stable Limit on (me- step size Numerical dispersion problems Small (me steps or high- order schemes Boundary

10 HPC Trends in Seismics In spite of their spectacular performance, multicores, GPUs and FPGAs obey to a restrictive programming paradigm, data parallelism or flow parallelism, that makes difficult the numerical solution of many algebraic problems...some new thinking must be done on problem formulations: change the physical paradigm!!!

parallelism, that makes difficult the numerical solution of many algebraic problems.

11 Problem formaliza,on The path to Seismic Software Development

12 The path to Seismic Software Development Problem formaliza,on Finding concurrency

13 The path to Seismic Software Development Problem formaliza,on Finding concurrency Algorithm structure and data- parallel encoding

14 The path to Seismic Software Development Problem formaliza,on Finding concurrency Industrial sokware as a service Algorithm structure and data- parallel encoding

15 Time Imaging: a Data-driven Solution xd x0 t0 td ZO- stacked data Time migrated data

16 Time Imaging: a Data-driven Solution xd x0 t0 td...this is a successful example where the use of a PDE solubon to image the acousbc medium is no longer necessary!!! ZO- stacked data Time migrated data

17 Conclusion When imaging may be recast as a large collec,on of local op,miza,on problems, its solu,on perfectly fits the restric,ve parallelism paradigm imposed by mul,cores, GPUs and dataflow engines, thus taking fully advantage of their spectacular performance Tacit knowledge is some,mes difficult to grasp and transfer but team work with HW manufacturers may provide great results

dataflow engines, thus taking fully advantage of their spectacular performance Tacit knowledge

18 ...looking for a funding industrial partner!!! Conclusion

M. Lejeune, C. López, V. Gestí, B. Tomás, A. Korzynska, A. Roso, C. Callau, R. Bosch, J. Baucells, J. Jaén. Digital image analysis

A multistep image analysis method to increase automated identification efficiency in immunohistochemical nuclear markers with a high background level M. Lejeune, C. López, V. Gestí, B. Tomás, A. Korzynska,