8 dome results · dome project: dome research streams… 1. algorithms & machines 2. access...
TRANSCRIPT
DOME
DOME: Results.
www.dome-exascale.nl
Prof. Dr. Ton Engbersen,IBM Research Laboratory – Zurich, SwitzerlandScientific Director ASTRON-IBM Center for Exascale Technology, NetherlandsProf. Data Science Engineering, Rijksuniversiteit Groningen, NetherlandsMember IBM Academy of Technology
DOMEDOME: What is it?
IBM-NL
IBM-ZRLASTRON
€ €
€ €
DOME
DOME: Why?
3
Platformen
Photonics
Modelling
Algorithms
Storage solutions
Access patterns
Dome Project:
SystemAnalysis
Data&StreamingSustainable(Green)Computing Nanophotonics
Computing Transport Storage
Modelling
- Nanophotonics- Real-TimeCommunications- NewAlgorithms
- Microservers- Accelerators
- AccessPatterns
ResearchStreams…
…aremappedtoresearchprojects:
…plusanopenuserplatform:Userplatform
- Studentprojects- Events- ResearchCollaboration
5-year Research Project: 76 IBM PY (32 in NL); 50 ASTRON PY
Impact of DOME on SKA:
DOME Symposium 18-19 Mei 2017
5
SCIENCE REGIONAL CENTRE
SRC
MicroserversPhotonics
AcceleratorsAlgorithms
Data access patterns
Modeling
Real time communications
Impact of DOME on SKA:
DOME Symposium 18-19 Mei 2017
6
SCIENCE REGIONAL CENTRE
SRC
MicroserversPhotonics
AcceleratorsAlgorithms
Data access patterns
Modeling
Real time communications
Dome Project: DOME
ResearchStreams…
1. Algorithms&Machines
2. AccessPatterns
3. Nanophotonics
4. Microserver
5. Accelerators
6. NewAlgorithms
7. Real-TimeCommunications
5-year Research Project: 76 IBM PY (32 in NL); 50 ASTRON PY
PISAPlatform-Independent
Software Analysis
workload properties ExaBounds system results
ExtrAX
Modeling – Tool Flow
Goal: Create a holistic design-space exploration tool to overcome fundamental technology limits in data centers, servers and exascalesystems by use of a novel formal method that captures first principles in form of equations compounded with boundary conditions (power, required throughput, I/O, technology parameters, architecture).
P1 – Modeling
SKA algorithms analysis→ compute requirements→ power analysis: power budgets hard to meet → power-efficient chip architectures (SKA signal-processing front-end)
+ ExaBounds (SKA back-end: SDP)
Bottom-line:
Impact of DOME on SKA:
DOME Symposium 18-19 Mei 2017
10
SCIENCE REGIONAL CENTRE
SRC
MicroserversPhotonics
AcceleratorsAlgorithms
Data access patterns
Modeling
Real time communications
11 Sep 10 2013 IBM Confidential
Access patterns: Data units assignment example (1000 100GB chunks)
Budget = $23,000Mean Response = 0.27 sec
Budget = $35,000Mean Response = 0.0017 sec
l Red: SSDl Green: HDDl Blue: Tape
P2 – Access Patterns
1. Fully analytical model for multi-tier storage system cost/performance calculation:• Used for APERTIF Long Term Archive • Basis for Regional science datacenters for SKA project.
2. Analytical model for performance analysis of predictive caching:• Used for LOFAR Long Term Archive at this momen
• “Cognitive Storage” project.
Bottom-line:
Impact of DOME on SKA:
DOME Symposium 18-19 Mei 2017
13
SCIENCE REGIONAL CENTRE
SRC
MicroserversPhotonics
AcceleratorsAlgorithms
Data access patterns
Modeling
Real time communications
Photonics – Overview
Sky Signal
Antenna LNA
ADCADCADCADC
Transport-links Beamforming & Filtering
BFCorrelator/CSP
Signal processing
few to 10 km
Optical Beamforming
Antenna(omnidirectional)
Analog Fiber Links
Sky Signal
Antenna-Element
LNA
LaserSky Signal
Antenna-Element
LNA
Laser
Photonics Beamforming
fiber cable PD
Amplifier
Sky Signal
Antenna-Element
LNA
Laser
Beamforming = Reducing Data: - Electronic/Digital Beamforming «in» the dish= Reducing # cables risks to interfere with signal reception (RFI)
- Optical Beamforming ... does NOT!!
Beamformer Requirements:Frequency Range: ~3 GHzArray Dimensions: 10 x 11 ElementsElement Spacing: 21 mmArray Size: 22.4 x 22.4 cmBeam-Steering: +/- 30º
PhotonicsBeamformer
15 Currently being measured
Impact of DOME on SKA:
DOME Symposium 18-19 Mei 2017
16
SCIENCE REGIONAL CENTRE
SRC
MicroserversPhotonics
AcceleratorsAlgorithms
Data access patterns
Modeling
Real time communications
Microserver - DefinitionµServer:
The integration of an entire server node motherboard*
into a single microchip except DRAM, Nor-boot flash and power conversion logic.
305mm
245mm
139mm x55mm x 7mm
This does NOT imply low performance!
* no graphics
138mm x67mm x 6.9mm
iPhone-6
Ultimate System: 2U rack unit
19” 2U Chassis w/ Combined Cooling & Power128 compute node boards
1536 cores / 3072 Threads6 TB DRAM1.28Tbps Ethernet (@40Gbps)à Datacenter-in-a-box
40% more performance @ 70% of node level energy consumption à 2x more operations per Watt
4.4: Energy-Efficient Microserver Based on a 12-Core 1.8GHz 188K-CoreMark 28nm Bulk CMOS 64b SoCfor Big-Data Applications with 159GB/s/L Memory Bandwidth System Density
2015 IEEE International Solid-State Circuits Conference
Followed through: ILA microservers start-up
Impact of DOME on SKA:
DOME Symposium 18-19 Mei 2017
19
SCIENCE REGIONAL CENTRE
SRC
MicroserversPhotonics
AcceleratorsAlgorithms
Data access patterns
Modeling
Real time communications
Image-Domain griddingExplore inherent parallelism in gridding...?
Convolution in Fourier Domain = multiplication in image domain
• ~32 x 32 subgrids
• Local memory• parallelism
subgrid
subgrid
It works !
Status: • On GPU (!) 25x faster then LOFAR (CPU) imager• Imager for LOFAR
Image courtesy of Bas van der Tol
Impact of DOME on SKA:
DOME Symposium 18-19 Mei 2017
21
SCIENCE REGIONAL CENTRE
SRC
MicroserversPhotonics
AcceleratorsAlgorithms
Data access patterns
Modeling
Real time communications
22 Sep 10 2013 IBM Confidential
Basic Algorithms:
Impact of DOME on SKA:
DOME Symposium 18-19 Mei 2017
23
SCIENCE REGIONAL CENTRE
SRC
MicroserversPhotonics
AcceleratorsAlgorithms
Data access patterns
Modeling
Real time communications
RDMA
ScientificDataProcessing
DataCorrelation
3
21
Efficiency of data transport1. Power consumption of data transfer and processing2. Efficient data transfer to data processing entity3. Intermediate ingress data buffering at proc. Entity4. Use of Flash as ingress data buffer.5. Underlying technologies for Spark.
1. 7 Very active workstreams with focus on SKA• 130+ Scientific (peer reviewed) publications since start, many more accepted & in
preparation,
• DOME impact on SDP – SKA1:• Performance modelling• SDP (and SKA) architecture – re-baselining• Image Domain gridding on accelerators• Microserver as data center in a box• RDMA for energy efficient data transport• Near-memory computing• Fundamental algorithm research• Optical transmission antennae – CSP/station-processing• Balance new vs. existing technologies• Technology vision• Energy Focus
• DOME impact on SKA-2:• ASIC’s, DSP• Nanophotonics & optical beamforming• Algorithms for beamforming and calibration• RDMA• Microservers/microdata centers