High Performance Computing for spacestream · AI DNN training? Quantum problems solving? (i.e....
Transcript of High Performance Computing for spacestream · AI DNN training? Quantum problems solving? (i.e....
High Performance Computing
for spacestream.
The future of EO!
ESA Workshop
Quantum Processing of Big Data
from quantum computing to earth observation
Rome – Italy – 3rd of April 2019
Cristoforo Abbattista
2
Planetek Group
3
Planetek Activities
Earth Observations
Environmental Monitoring
Spatial Data Infrastructure
Urban Planning
Oil & Gas - Renewable Energy
Infrastructure Monitoring
Ground Segment
On board Software
Cosmic Exploration
4
Planetek in the Space Value Chain
PLANNING &TASKING
5
THE SPACESTREAM
THE RIGHT INFORMATION
AT THE RIGHT TIME , IN THE RIGHT PLACE
6
The New SpaceStream
Networks of “heterogeneous, distributed Ground Segments”
Swarms of ”cooperative and competitive Space Agents”
UpStream and DownStream mixed in the FOG
of the Continuous SpaceStream
to reach the goals of
Responsiveness, Reactivity and Low Latency
7
ESA Vision
8
…let’s start from the end
9
10
11 © kfor.com
Displacement
12 © kfor.com
Displacement
13
Challenge
Identifying subsidence before it becomes critical
14
15
16
Rheticus® Marine Chlorophyll
17
Rheticus® AquacultureAquaculture
18
Rheticus® AquacultureAquaculture
19
...very demanding activities to
perform
GPU4Space Library
Multi OS: MS Windows, Linux, MacOS
Multi Processor: Intel, AMD, ARM, Sparc
Multi GPU: Intel, AMD, NVIDIA, ARM-Mali,
Imagination-PowerVR
Multi GPU-API: OpenGL, OpenCL, CUDA,
Vulkan
Multi-platform: CPU/GRID/GPU
Data Oriented: Tensor and hierarchical
Tensor (HDF5, NetCDF, FITS)
Constrained Memory Access: Protection
from SEU (Single-Event Upset) like buffer
overflow/underflow and not allowed access
to memory
Service Locator Pattern: to extend at run
time the execution programme without
recompiling
21
SAR PDGS workflow30 seconds of acquisition need 60 minutes of processing
ACQ
Storage
Archive
&
Catalogue
PM
Storage
Archive
&
Catalogue
CM USERS
Acquisitionand
Deformatting
Level X Processing Distribution
SAR Standard products
22
Persistent Scatterer Processing chain
Main steps
Coregistration
Differential interferogram
generation
Estimation and removal of
phase artifacts (orbital and
atmospheric)
PS map generation
23
More expensive steps CPU time
(hours)
GPU
acceleration
SLC stack coregistration 500 5x
PS candidates detection 2.400 24x
Differential velocity estimation of PS
candidates
800
Atmospheric contribution estimation
and removal
3.300 10x
TOTAL 8.000 h
333 days
6x (1.300 h
54 days)
PS chain profiling (Cosmo SKYMED products)
For a stack of 70 Stripmap images (400 Mpx each)
24
Automatic Geocoding
10 images (~ 2 Gpx)Processing Speed of about 100 Mpx/h
Operation Time
Pre-Processing 2 h
Matching 12 h
Orthorectification 1,5 h
25
…not only
Earth Observation
26
The BeyondPLANCK project
27
The BeyondPLANCK project
The likelihood function to solve3000
3000
105 likelihood evaluations
full exploration of the cosmological parameters space
CPU 8core,
16thread
CPU 16core,
32thread
CPU 16core,
32thread
GPU 1
(64 bit floating point)
GPU 2
(64 bit floating
point)
Once ~2000 ms ~700 ms ~200 ms ~1300 ms ~4500 ms
105 23 days 8,1 days 2,3 days 15 days 52 days
Central Processing Unit Graphics Processing Unit
28
Satellite Health Monitoring
Thousands of HKTM each secondto correlate & investigate!
29
Quantum Imaging Sensor
CORRELATED
Light Source
SENSOR 1: retrieves the
direction of light rays
SENSOR 2: retrieves the
"ghost“ image of the scene
Thousands of images to be cross-correlated!
30
Is Quantum Computing also a threat?
Blockchain 4 Spacestream
CyberSecurity
31
Conclusions
We need fast computing.
Also/even as a service!
Should we deploy DIAS close to QC facilities?
Where QC can give its best?
Simulations?
Direct Model computation?
Model inversion?
AI DNN training?
Quantum problems solving? (i.e. Quantum Imaging)
Can be quantum4EO library a GOAL?
Is Quantum also a threat?
32
Thank you for your attention
Cristoforo Abbattista
Head of Planetek SpaceStream SBU