Upload
others
View
8
Download
0
Embed Size (px)
Citation preview
CoreNeuron : Morphologically Detailed Neuron Simulations
Building, Simulating and Optimizing Large Neuron Networks on GPUs
Pramod Kumbhar, Michael Hines& Blue Brain HPC Team7th April 2016, GTC
Understanding Brain..
2013
BRAIN Initiative
2013
Human Brain Project
2014Brain/MINDS
Better understanding of brain
2007
Izhikevic: brain scale simulation on cluster
1 million cells
2009
IBM: Cat’s brain scale simulation on BG-P
1.5 billion cells
See source1
Brain Simulations
Point Neurons Morphologically Detailed Neurons
Molecular Level
See source3 See source4
Blue Brain Project, EPFL
Reverse Engineering Brain (10000 feet view)
See source5
See source6
Reconstruction Workflow
• 15+ years of Experiments
• 14,000 Neurons recorded
and labeled
• 2,052 Classified neurons
• 1,009 Reconstructed
neurons
• 2,000+ Ion channel
recordings
• 4,000+ Electrical recordings
of single neurons
• 5,000+ Synaptic recordings
of pairs of neurons
Markram et al. 2015, Cell
Modeling Neuron
See source8
See source7
HH, 1952
• Prominent components of nervous system
• More than 300 ion channels
Ion Channels
Biologist view:compartment model
Every channel is a compute kernel, no single hotspot!
NMODL: Source to Source Compiler
LexicalAnalyzer
SyntaxAnalyzer
SemanticAnalyzer
~ IntermediateRepresentation
tokens
parse tree
parse tree
OpenACCC Cuda Cyme:SIMDDSL
backends
NMODL
PortablePerformance
wrap OpenACC and vectorisation hints related pragmas
auto-generated kernel
OpenACC API’s to copy the complex data structure
OpenACC Kernels
User defined DS: Major challenge for many application
AoS/SoA, Vectorisation, Memory Coalescing etc..
bksub: for(i = x; i < nodes; i++) {
rhs[i] -= b[i] * rhs[ parent[i] ]rhs[i] /= d[i]
}
GPU: Cell level Parallelism (some kernels)
node : 1 2 3 4 parent[i]: 0 1 2 3
step 0
Step 1
Step 2
Step 3
node : 6 7 8 9 parent[i]: 5 6 7 8
node : 156 157 158 159 parent[i]: 155 156 157 158
thread 0
thread 1
thread 31
Memory addresses0
warp
Stride Depth of Tree
0
1
2
3
4
-1
0
1
2
3
parent_indexnode_index
5
6
7
8
9
-1
5
6
7
8
0
2
4
6
8
-1
0
1
2
3
parent_indexnode_index
1
3
5
7
9
-1
5
6
7
8
0
1
2
3
4
5
6
7
8
9
-1
-1
0
1
2
3
4
5
6
7
all cells root
Cell 0 and Cell 1 nodes interleaved
Roots parent
Cell 0 parent
Cell 1 parent
Nodes Parents
Cell Interleaving
Permutations: ions, synapses, areas, point processes..
Memory addresses0
warp
Spike ExchangeA
20 mV200 ms
bAC
dNAC
cSTUT
bSTUT
cIR
bIR
cAD
cAC
dSTUTcNAC
bNAC
e-ty
pe fr
actio
n fo
r eac
h m
-type
m-ty
pe
e-type
100%
80
60
40
20
0
C
cAC
cNAC
bNAC
dNACB
bAC
cIR
37% 31% 5%
25% 1.5% 1.5%
Figure 4. Markram et al.
e-types me-types me-combinationsL1DAC
L1NGC-DA
L1NGC-SA
L1HAC
L1LAC
L1SAC
L23PC
L23MC
L23BTC
L23DBC
L23BP
L23NGC
L23LBC
L23NBC
L23SBC
L23ChC
L4PCL4SP
L4SS
L4MC
L4BTC
L4DBC
L4BP
L4NGC
L4LBC
L4NBC
L4SBC
L4ChC
L5TTPC1
L5TTPC2
L5UTPC
L5STPC
L5MC
L5BTC
L5DBC
L5BP
L5NGC
L5LBC
L5NBC
L5SBC
L5ChC
L6TPCL1
L6TPCL4
L6UTPC
L6IPC
L6BPC
L6MC
L6BTC
L6DBC
L6BP
L6NGC
L6LBC
L6NBC
L6SBC
L6ChC
cAC
bAC
cNAC
bNAC
dNAC
cSTU
TbS
TUT
dSTU
T cIR bIR cAD
L23 NBC(burst Accommodating)
(continuous Non-accommodating)
(burst Non-accommodating)
(continuous Accommodating)
(delayed Non-accommodating)
(continuous Stuttering)
(burst Stuttering)
(delayed Stuttering)
(continuous Irregular)
(burst Irregular)
(continuous Adapting)
• Electrical diversity: 11 e-types; 207 me-types
• Number of connections increases exponentially
• Different types of events
NetCon List Buffering Mechanism
Compute engine of NEURON simulator
Being developed for large scale simulations (28 racks BG-Q)
Ion Channels: ~ 85% time
Linear Algebra: ~ 5-7%
Spike Exchange: 7-10%
CoreNeuron
Timeline
loadCPU
GPUcurrent solve state
dt
copy initialize
setup
threshold
queue MPI queue
mindelay
Toolchain
Soma
Compartment
hhpas
pas
pas
pas
pas
Simple model…
larger cells
65536 3072 1024
Varying # : Rings, Cells, Branches, Compartments
Model A Model B Model C Model D Model E
Performance
• K20x vs 8-core Xeon• Cray (OpenACC)• Cuda 7• No “hand” tuning yet• Optimized CPU code with
vectorization
Real World Models
Ion channels are 4-8x faster in all models!
Kernels with cell level parallelism, low occupancy!
393216 65536 4096 3072
• 8-core Xeon / 8 MPIs• BG-Q Node / 32-64 threads• Xeon Phi 61 core @ 1.23 GHz,
180-240 threads• K20X GPU• Optimized Xeon/MIC code with
vectorization (XLC issue)
Performance
Varying # : Rings, Cells, Branches, Compartments
Cell Interleaving and Exposing Parallelism
HomogenousHeterogeneous
Ideal
ill - suited
How much parallelism? How much imbalance?
Morphological diversity challenge
“Morphology Aware Scheduling of Kernels using Isomorphic Subtrees”
Resource
Reconstruction and Simulationof Neocortical MicrocircuitryHenry Markram,1,2,19,* Eilif Muller,1,19 Srikanth Ramaswamy,1,19 Michael W. Reimann,1,19 Marwan Abdellah,1
Carlos Aguado Sanchez,1 Anastasia Ailamaki,16 Lidia Alonso-Nanclares,6,7 Nicolas Antille,1 Selim Arsever,1
Guy Antoine Atenekeng Kahou,1 Thomas K. Berger,2 Ahmet Bilgili,1 Nenad Buncic,1 Athanassia Chalimourda,1
Giuseppe Chindemi,1 Jean-Denis Courcol,1 Fabien Delalondre,1 Vincent Delattre,2 Shaul Druckmann,4,5
Raphael Dumusc,1 James Dynes,1 Stefan Eilemann,1 Eyal Gal,4 Michael Emiel Gevaert,1 Jean-Pierre Ghobril,2
Albert Gidon,3 Joe W. Graham,1 Anirudh Gupta,2 Valentin Haenel,1 Etay Hay,3,4 Thomas Heinis,1,16,17 Juan B. Hernando,8
Michael Hines,12 Lida Kanari,1 Daniel Keller,1 John Kenyon,1 Georges Khazen,1 Yihwa Kim,1 James G. King,1
Zoltan Kisvarday,13 Pramod Kumbhar,1 Sebastien Lasserre,1,15 Jean-Vincent Le Be,2 Bruno R.C. Magalhaes,1
Angel Merchan-Perez,6,7 Julie Meystre,2 Benjamin Roy Morrice,1 Jeffrey Muller,1 Alberto Munoz-Cespedes,6,7
Shruti Muralidhar,2 Keerthan Muthurasa,1 Daniel Nachbaur,1 Taylor H. Newton,1 Max Nolte,1 Aleksandr Ovcharenko,1
Juan Palacios,1 Luis Pastor,9 Rodrigo Perin,2 Rajnish Ranjan,1,2 Imad Riachi,1 Jose-Rodrigo Rodrıguez,6,7
Juan Luis Riquelme,1 Christian Rossert,1 Konstantinos Sfyrakis,1 Ying Shi,1,2 Julian C. Shillcock,1 Gilad Silberberg,18
Ricardo Silva,1 Farhan Tauheed,1,16 Martin Telefont,1 Maria Toledo-Rodriguez,14 Thomas Trankler,1 Werner Van Geit,1
Jafet Villafranca Dıaz,1 Richard Walker,1 Yun Wang,10,11 Stefano M. Zaninetta,1 Javier DeFelipe,6,7,20 Sean L. Hill,1,20
Idan Segev,3,4,20 and Felix Schurmann1,201Blue Brain Project, Ecole polytechnique federale de Lausanne (EPFL) Biotech Campus, 1202 Geneva, Switzerland2Laboratory of Neural Microcircuitry, Brain Mind Institute, EPFL, 1015 Lausanne, Switzerland3Department of Neurobiology, Alexander Silberman Institute of Life Sciences, The Hebrew University of Jerusalem, Jerusalem 91904, Israel4The Edmond and Lily Safra Center for Brain Sciences, The Hebrew University of Jerusalem, Jerusalem 91904, Israel5Janelia Farm Research Campus, Howard Hughes Medical Institute, Ashburn, VA 20147, USA6Laboratorio Cajal de Circuitos Corticales, Centro de Tecnologıa Biomedica, Universidad Politecnica de Madrid, 28223 Madrid, Spain7Instituto Cajal (CSIC) and CIBERNED, 28002 Madrid, Spain8CeSViMa, Centro de Supercomputacion y Visualizacion de Madrid, Universidad Politecnica de Madrid, 28223 Madrid, Spain9Modeling and Virtual Reality Group, Universidad Rey Juan Carlos, 28933 Mostoles, Madrid, Spain10Key Laboratory of Visual Science and National Ministry of Health, School of Optometry and Opthalmology, Wenzhou Medical College,Wenzhou 325003, China11Caritas St. Elizabeth’s Medical Center, Genesys Research Institute, Tufts University, Boston, MA 02111, USA12Department of Neurobiology, Yale University, New Haven, CT 06510 USA13MTA-Debreceni Egyetem, Neuroscience Research Group, 4032 Debrecen, Hungary14School of Life Sciences, University of Nottingham, Nottingham NG7 2UH, United Kingdom15Laboratoire d’informatique et de visualisation, EPFL, 1015 Lausanne, Switzerland16Data-Intensive Applications and Systems Lab, EPFL, 1015 Lausanne, Switzerland17Imperial College London, London SW7 2AZ, UK18Department of Neuroscience, Karolinska Institutet, Stockholm 17177, Sweden19Co-first author20Co-senior author*Correspondence: [email protected]://dx.doi.org/10.1016/j.cell.2015.09.029
SUMMARY
We present a first-draft digital reconstruction of themicrocircuitry of somatosensory cortex of juvenilerat. The reconstruction uses cellular and synapticorganizing principles to algorithmically reconstructdetailed anatomy and physiology from sparse experi-mental data. An objective anatomical method definesa neocortical volume of 0.29 ± 0.01 mm3 containing!31,000 neurons, and patch-clamp studies identify55 layer-specific morphological and 207 morpho-electrical neuron subtypes. When digitally recon-structed neurons are positioned in the volume andsynapse formation is restricted to biological boutondensities and numbers of synapses per connection,
their overlapping arbors form !8 million connectionswith !37 million synapses. Simulations reproducean array of in vitro and in vivo experiments withoutparameter tuning. Additionally, we find a spectrumofnetworkstateswithasharp transition fromsynchro-nous to asynchronous activity, modulated by physio-logical mechanisms. The spectrum of network states,dynamically reconfigured around this transition, sup-ports diverse information processing strategies.
INTRODUCTION
Since Santiago Ramon y Cajal’s seminal work on the neocortex(DeFelipe and Jones, 1988; Ramon y Cajal, 1909, 1911), a vastnumber of studies have attempted to unravel its multiple levels
456 Cell 163, 456–492, October 8, 2015 ª2015 Elsevier Inc.
Resource
Reconstruction and Simulation of NeocorticalMicrocircuitry
Graphical Abstract
Highlightsd The Blue Brain Project digitally reconstructs and simulates a
part of neocortex
d Interdependencies allow dense in silico reconstruction from
sparse experimental data
d Simulations reproduce in vitro and in vivo experiments
without parameter tuning
d The neocortex reconfigures to support diverse information
processing strategies
AuthorsHenry Markram, Eilif Muller,
Srikanth Ramaswamy,
Michael W. Reimann, ..., Javier DeFelipe,
Sean L. Hill, Idan Segev, Felix Schurmann
In BriefA digital reconstruction and simulation of
the anatomy and physiology of
neocortical microcircuitry reproduces an
array of in vitro and in vivo experiments
without parameter tuning and suggests
that cellular and synaptic mechanisms
can dynamically reconfigure the state of
the network to support diverse
information processing strategies.
Markram et al., 2015, Cell 163, 456–492October 8, 2015 ª2015 Elsevier Inc.http://dx.doi.org/10.1016/j.cell.2015.09.029
THANK YOU!
Explanatory Graphic Sources
• source 1: dgallery.s3.amazonaws.com
• source 2: clipartbest.com/cliparts, scaryforkids.com, geniusawakening.com
• source 3: developer.humanbrainproject.eu
• source 4: nature.com
• source 5: deviantart.net, squarespace.com
• source 6: lcn.epfl.ch
• source 7: nature.com
• source 8: genesis-sim.org
low occupancy!
Model A Model B Model C Model D Model E
larger cells
65536 2048 1024
Backup