Real-time Monitoring of Large Scientific Simulations D. E. Laney - - PowerPoint PPT Presentation

▶
real time monitoring of large scientific simulations d e
SMART_READER_LITE
LIVE PREVIEW

Real-time Monitoring of Large Scientific Simulations D. E. Laney - - PowerPoint PPT Presentation

Jun 16, 2023 330 likes •550 views

Real-time Monitoring of Large Scientific Simulations D. E. Laney V. Pascucci, R. J. Frank, G. Scorzelli, L. Linsen, B. Hamann, F. Gygi This work was performed under the auspices of the U.S. Department of Energy by the University of California,

slide-1
SLIDE 1

Real-time Monitoring of Large Scientific Simulations

  • D. E. Laney
  • V. Pascucci, R. J. Frank, G. Scorzelli,
  • L. Linsen, B. Hamann, F. Gygi

This work was performed under the auspices of the U.S. Department of Energy by the University of California, Lawrence Livermore National Laboratory under Contract No. W-7405-Eng-48.

slide-2
SLIDE 2

We would like to enable real-time monitoring of a running simulation with a desktop workstation.

Turbulent mixing: 2048 ^ 3 Data post-processed for visualization

slide-3
SLIDE 3

Two ideas are combined in our system.

  • Multiresolution re-ordering of the data
  • Real-time streaming of the reordered data.
  • Monitoring via visualization of data or

derived-data.

slide-4
SLIDE 4

A data streaming framework must utilize a heterogeneous computational environment.

Data Servers Data Sources Data Client (Desktop Display) Data Server Data Client (Powerwall Display)

slide-5
SLIDE 5

A brute force approach: send the data in the order it occurs in the simulation.

  • Disadvantages
  • Bad cache coherence will slow down lower

dimensional queries (slices, isosurfaces).

  • Stopping the stream early results in an unusable

partial data set.

  • Requires preprocessing and delay if scalable

visualization is required.

slide-6
SLIDE 6

We propose a new multiresolution streaming system.

  • Light weight:
  • Only requires init() and send() calls.
  • Not a problem solving framework
  • Cache-oblivious out-of-core processing and

data streaming.

  • Simple, scalable infrastructure.
slide-7
SLIDE 7

Our approach enables a coarse-to-fine construction of multi-resolution models.

Jeep run (Francois Gygi) of 64 carbon atoms, tiled to 256^3 (final iteration). Computation on AIX system, reordered data cached on SGI server, viewed on my workstation (200MHz SGI).

slide-8
SLIDE 8

We exploit the correlation of bin/quad/oct- trees with the Lebesgue space-filling curves.

The Lebesgue curve is also known as Z-order, Morton, …. Curve.

Special case of the general definition introduced by Guiseppe Peano in 1890.

slide-9
SLIDE 9

Progressive data streams enable multiresolution visualization.

(a) (b) (c) (d) coarse data new level data

  • On the fly

hierarchical Z-

  • rdering

– Embedded preprocessing – Fast computation

  • f hierarchical

index – Stream can be truncated

slide-10
SLIDE 10

Our cache oblivious data layout optimizes visualization queries.

  • Data is laid out by

level and broken into blocks.

  • Simple load

balancing on visualization side.

  • Fast slicing,

isosurfacing, and volume rendering.

slide-11
SLIDE 11

The data layout generates increasingly local coverage.

Distribution in the grid of each constant size block of data 0 1 2 3 4 5 6 7 8

slide-12
SLIDE 12

We implement immediate streaming of simulation data to the storage location.

Simulation Visualization System

Data Server 1 Data Source 1 Data Source 2 Data Source N Data Server 2 Data Server M

slide-13
SLIDE 13

We exploit the data decomposition used in the simulation.

B1 B2 B3 B4 B9 B10 B11 B12 B5 B6 B7 B8 B13 B14 B15 B16

  • Takes advantage of

simulation load balancing

  • One Data Source per

compute node

  • Each compute node

connected to all Data Servers

slide-14
SLIDE 14

Data Servers are buffering and caching components.

Data Server 1 Data Source 1 Data Source 2 Data Source N Data Server 2 Data Server M

  • Cache data and respond to visualization

queries

  • Filter and buffer data to produce multi-

layer configurations

slide-15
SLIDE 15

We achieve load balancing with a static data decomposition.

  • No contention
  • No data duplication
  • Streaming:

– D = I/d mod N

– D is Data server, I is HZ index of sample, d is chunk size

  • Visualization queries:

– Choose number of Data Servers wisely

slide-16
SLIDE 16

The load balancing of visualization queries depends on the total number of data servers.

0.5 1 1.5 2 2.5 3 3.5 4 4.5 10 20 30 40 50 60 70 80 90 100

Max Load / Ideal Load Number of Data Servers Optimal

slide-17
SLIDE 17

The measured streaming time is small compared to the compute time.

Number of Data Sources 1 8 64 Number of Data Servers 1 3 3 Total Domain Size 1283 2563 5123 Equivalent simulation time/time- step 270s 736s 4224s Send Time 5.31s 7.03s 38.6s

slide-18
SLIDE 18

A progressive data stream enables visualization on desktop workstations.

Progressive refinement (left to right) of the volume rendering of the electron density distribution.

slide-19
SLIDE 19

Data servers handle localized queries.

slide-20
SLIDE 20

On a desktop machine you can explore the finest resolution data available.

~ 8 billion values

slide-21
SLIDE 21

Future Work

  • Extend to non-power-of-two grid sizes
  • Extend to more general mesh structures
  • Include fast compression between data

source and data server

  • Enable multiple layers of data servers to

handle larger simulation sizes.