A CUDA-Based 3D Kinetic Model for Space Plasma Physics Shahab - - PowerPoint PPT Presentation

A CUDA-Based 3D Kinetic Model for Space Plasma Physics

Shahab Fatemi and Andrew R. Poppe

Space Sciences Laboratory UC Berkeley, CA, USA

GPU T echnology Conference April 7, 2016

Image courtesy of S. Saarloos

2

 Plasma is the fourth state of matter.  Highly ionized gas, often >100,000 Kelvin.  Mainly consists of electrons and ions (charged particles).  Solar wind (supersonic fmow of plasma from the Sun)

typical speed near Earth: ~400 km/s (~895,000 mi/h)

[NASA]

3

Solar wind plasma interacts with difgerent objects in difgerent ways.

Mars Earth

[Artistic work: NASA/GSFC]

4

[ C a n a d a . P h

• t
• g

r a p h e r : D a n i e l J . C

• x

/ C

• r

b i s ]

A visual proof for solar wind interaction with the Earth: Northern lights (aurora) Aurora on Jupiter, but not visible by eye (Ultra Violet).

[John Clarke (University of Michigan)]

5

 Solar wind interaction with the Earth impacts our daily life.  Astronauts, and spacecrafts in space, and their safety.  Fundamental information on the history of planetary evolution

Mars has lost most of its atmosphere through interaction with solar wind plasma.

 Space weather monitors solar activity.  Solar wind has been interacting with the Earth over billions of

years, since the Earth was born.

Why are these important?

6

[STEL, Nagoya University]

In-situ observations Lab experiments Modeling/Simulations

[NASA/New Horizons] [Vacuum chamber/University of Colorado]

Modeling techniques

7

Fluid Kinetic

Magnetohydrodynamic (MHD) For global scale simulations Boltzmann-Vlasov Particle-in-cell (PIC) For small scale simulations Hybrid (Kinetic-Fluid) Similar to PIC models For medium scale simulations

Motivation

8

Major problems in HPC in space and plasma physics:

 Network latency (communication)  Processor load imbalance (source and loss processes)  High cost to maintain super computers  CPU performance

HPC using CPUs in space and plasma modeling has reached a state that cannot satisfy all our needs. We need to use new technologies. We need GPUs!

Kinetic-particle method

9

Similar to n-body problems. Move charged particles. Solve electromagnetic (wave) equations.

Our GPU-based kinetic model

10

In our implementation, we use Single-CPU Single-GPU

Plasma motion

11

Concept of plasma motion is difgerent than neutral atoms and thermal gases.

Particle mapping to grid

12

Performance (naïve implementation)

13

Grid size: 100x100x100 Particles per cell: 64

(total: 64M particles)

Particle mapping to grid

14

Particles are marked in every block of particles. A thread-block is assigned to a block of particles. Particles are sorted as they move between grid cells.

Comparison

15

Atomic operations are not all that bad!

Function Naïve (ms) Advance (ms) Speed up Particle motion 46.6 38.4 ~1.20 Particle mapping 261.7 189.6 ~1.38 Field mapping 187.3 84.9 ~2.20 Grid size: 100x100x100 Particles per cell: 64

(total: 64M particles)

Application 1

16

Plasma interaction with Ganymede (the largest moon in our solar system)

Earth Ganymede Moon

Fair Comparison

17

A fair comparison is always a challenge! Be fair and keep developing! We made two identical simulation runs (identical grid cells, particle number, time steps) Run #1) 288 Intel processors (6 nodes x 48 CPUs) Without GPUs Run #2) 1 Intel processor + single Titan X GPU.

Speed up: Cost not included 158/124 =1.27 Cost included 158*60/124 = 76 Model Run time (h) Cost ($) CPU-based ~158 ~(60K?) GPU-based ~124 ~1K

Application 2

18

Real-time simulation of plasma interaction with Mars.

Mars Video here!

Summary

19

 The fjrst GPU based 3D kinetic model in space and plasma.  New algorithms introduced in our model.  We can now take a step forward to solve more complex problems.

Future work:

 Implement a multi-GPU model  Improvement in our algorithms

Field Solver

21

We use fjnite difgerence approximation to solve our electromagnetic equations.