high performance computing for neutron diffusion and ... · for neutron diffusion and transport...
TRANSCRIPT
High performance computingfor neutron diffusion and transport
equations
Horizon Maths 2012
Fondation Science Mathématiques de Paris
A.-M. Baudron, C. Calvin, J. Dubois,E. Jamelot, J.-J. Lautard, O. Mula-Hernandez
Commissariat à l’Énergie Atomique etaux Énergies Alternatives
Centre de Saclay
CEA/DEN/DANS/DM2S/SERMA/LLPR
December 20, 2012
Physical background MINOS solver MINARET solver Conclusions
Outline
Physical background
MINOS solver
MINARET solver
Conclusions
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 2 / 26
Physical background MINOS solver MINARET solver Conclusions
Top view of a nuclear reactor core (PWR 900MW)
After discretization: millions of unknowns. .
⇒ HPC to reduce volume of data and CPU time.
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 3 / 26
Physical background MINOS solver MINARET solver Conclusions
Fission chain reaction
Target nucleus
γ
(x, Ω, E)
NeutronFission product
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 4 / 26
Physical background MINOS solver MINARET solver Conclusions
Fission chain reaction
Target nucleus
γ
(x, Ω, E)
NeutronFission product N : neutron density,
ψ = vN : neutron angular flux,
J = Ωψ: neutron angular current.∂N∂t
= Production rate − Loss rate,
∂N
∂t=
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 4 / 26
Physical background MINOS solver MINARET solver Conclusions
Fission chain reaction
Target nucleus
γ
(x, Ω, E)
NeutronFission product N : neutron density,
ψ = vN : neutron angular flux,
J = Ωψ: neutron angular current.∂N∂t
= Production rate − Loss rate,
Σs: scattering cross section,
Σf : fission cross section,
Sext: external sources,
∂N
∂t=
∫
+∞
0
∫
S2Σs
(
x, (Ω,Ω′), E′→ E
)
ψ(x,Ω′, E
′)dΩ
′dE
′
+χ(E)
4π
(∫
+∞
0
ν(E′) Σf (x, E
′)
∫
S2ψ(x,Ω
′, E
′)dΩ
′dE
′
)
+Sext
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 4 / 26
Physical background MINOS solver MINARET solver Conclusions
Fission chain reaction
Target nucleus
γ
(x, Ω, E)
NeutronFission product N : neutron density,
ψ = vN : neutron angular flux,
J = Ωψ: neutron angular current.∂N∂t
= Production rate − Loss rate,
Σs: scattering cross section,
Σf : fission cross section,
Sext: external sources,
div J: leakage and streaming,
Σt: total cross section.
∂N
∂t=
∫
+∞
0
∫
S2Σs
(
x, (Ω,Ω′), E′→ E
)
ψ(x,Ω′, E
′)dΩ
′dE
′
+χ(E)
4π
(∫
+∞
0
ν(E′) Σf (x, E
′)
∫
S2ψ(x,Ω
′, E
′)dΩ
′dE
′
)
+Sext
−div J(x,Ω, E)−Σt(x, E)ψ(x,Ω, E).
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 4 / 26
Physical background MINOS solver MINARET solver Conclusions
Fission chain reaction
Target nucleus
γ
(x, Ω, E)
NeutronFission product N : neutron density,
ψ = vN : neutron angular flux,
J = Ωψ: neutron angular current.∂N∂t
= Production rate − Loss rate,
Σs: scattering cross section,
Σf : fission cross section,
Sext: external sources,
div J: leakage and streaming,
Σt: total cross section.
∫
+∞
0
∫
S2Σs
(
x, (Ω,Ω′), E′→ E
)
ψ(x,Ω′, E
′)dΩ
′dE
′
+χ(E)
4π
(∫
+∞
0
ν(E′) Σf (x, E
′)
∫
S2ψ(x,Ω
′, E
′)dΩ
′dE
′
)
=
div J(x,Ω, E)+Σt(x, E)ψ(x,Ω, E).
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 4 / 26
Physical background MINOS solver MINARET solver Conclusions
Criticality calculation: 6 variables
div J(x,Ω, E)+Σt(x, E)ψ(x,Ω, E)
=
∫ +∞
0
∫
S2
Σs
(
x, (Ω,Ω′), E′ → E)
ψ(x,Ω′, E′)dΩ′dE′
+1
λ
χ(E)
4π
(∫ +∞
0
ν(E′)Σf (x, E′)
∫
S2
ψ(x,Ω′, E′)dΩ′dE′
)
Physical solution: ψ ≥ 0, λ = keff | 1/keff = smaller EV.
Criticality:
keff < 1: production rate < loss rate, subcritical state,keff = 1: production rate = loss rate, critical state,keff > 1: production rate > loss rate, supercritical state.
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 5 / 26
Physical background MINOS solver MINARET solver Conclusions
Description of MINOS [Baudron, Lautard 2007]
• Criticality calculations, source problems, kinetics.• Cartesian or hexagonal mesh.• Energy E: multigroup theory.Eg ∈ [EG, EG−1] ∪ ... ∪ [E1, E0], EG < E0.
• Angular discretization Ω :Simplified spherical harmonics SPN (coupled diffusion equations).Diffusion.
• Space discretization x: Raviart-Thomas-Nédélec finite elements.Accurate computation of the neutron current and flux.
• C++, librairie boost, MPI.
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 6 / 26
Physical background MINOS solver MINARET solver Conclusions
SP1 multigroup equations
• φg(x) =∫ Eg−1
Eg
∫
S2ψ(x,Ω, E)dΩ dE,
• pg(x) =∫ Eg−1
Eg
∫
S2J(x,Ω, E)dΩ dE,
Solve in (pg, φg) |
σgr,1 p
g +grad φg =∑
g′ 6=g
σg′→gs,1 pg′
,
divpg +σgr,0 φ
g =∑
g′ 6=g
σg′→gs,0 φg
′
+1
λχg
G∑
g′=0
νg′σg′
f φg′
.
• σgr,0 and σg
r,1: Removal cross-sections of order 0 and 1,• σg
s,0 and σgs,1: Scattering cross-sections of order 0 and 1,
• νg: Nb of neutrons emitted by fission,
• σgf : Fission cross section, χg : Fission spectrum.
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 7 / 26
Physical background MINOS solver MINARET solver Conclusions
MINOS solver algorithm (criticality)
SPN solverEnergy sweeping
Outer iterations
of keffComputationSource
computation fission sourceUpdate of the
Acceleration of the outer iterations by means of Chebychev polynomials.The outer iterations lead the convergence.
More details in [Baudron, Lautard 2007].
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 8 / 26
Physical background MINOS solver MINARET solver Conclusions
Schwarz iterative method with optimized Robin interface conditions
[Lions (DD3) 1990, Guérin 2007, Nataf and Nier 1994].Let (p0
I , φ0I). ∀n ∈ N, solve in (pn+1
I , φn+1I ) such that :
σr,1 pn+1I +grad φn+1
I = QI in RI ,
divpn+1I +σr,0 φ
n+1I = SI in RI ,
φn+1I = 0 on ∂R∩ ∂RI ,
−pn+11 · n1 +α1φ
n+11 = pn
2 · n2 +α1φn2 on Γ,
−pn+12 · n2 +α2φ
n+12 = p
n(+1)1 · n1 +α2φ
n(+1)1 on Γ.
α1, α2 positive parameters. (pnI , φ
nI ) → (p, φ)|RI
with α1 = α2.
SPN case: more details in [Jamelot, Baudron, Lautard 2012].
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 9 / 26
Physical background MINOS solver MINARET solver Conclusions
Schwarz iterative method with optimized Robin interface conditions
[Lions (DD3) 1990, Guérin 2007, Nataf and Nier 1994].Let (p0
I , φ0I). ∀n ∈ N, solve in (pn+1
I , φn+1I ) such that :
σr,1 pn+1I +grad φn+1
I = QI in RI ,
divpn+1I +σr,0 φ
n+1I = SI in RI ,
φn+1I = 0 on ∂R∩ ∂RI ,
−pn+11 · n1 +α1φ
n+11 = pn
2 · n2 +α1φn2 on Γ,
−pn+12 · n2 +α2φ
n+12 = p
n(+1)1 · n1 +α2φ
n(+1)1 on Γ.
Other DD at EDF: Lagrange multiplier [Lathuilière 09].
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 9 / 26
Physical background MINOS solver MINARET solver Conclusions
MINOS solver algorithm with domain decomposition
Outer iterations
Subdomain j
Subdomain i
Exchange of the interfaceconditions if Γij 6= ∅
fission sourcecomputationSource
MPI calls
computationSource
Outer iterations
Exchange of the scalarproducts on the sources
Computationof keff
Computationof keff
Energy sweepingSPN solver
Energy sweepingSpatial solver
Update of thefission source
Update of the
The outer iterations lead the convergence. One DD iteration only.
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 10 / 26
Physical background MINOS solver MINARET solver Conclusions
Plate-fuel nuclear core.
SP3, 4 energy groups, RTN0,364× 364× 100 unit meshes: 425× 106 unknowns,εf = 5× 10−5.
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 11 / 26
Physical background MINOS solver MINARET solver Conclusions
Power maps
keff = 1.078065
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 12 / 26
Physical background MINOS solver MINARET solver Conclusions
Efficiency and accuracy tests (Titane, CCRT)
SP3, 4 energy groups, RTN0,364× 364× 100 unit meshes: 425× 106 unknowns.εf = 5× 10−5. Eff.= T1/(N × TN ) (%).
N (x, y, z) Nout Err. (10−5) CPU (s) Eff.1 (1, 1, 1) 644 0.0 7 165 100%2 (2, 1, 1) 651 0.0 4 175 86%4 (2, 2, 1) 642 0.0 2 290 78%8 (2, 2, 2) 654 0.0 1 285 70%16 (2, 2, 4) 650 0.0 639 70%32 (4, 4, 2) 657 0.4 319 70%64 (4, 4, 4) 659 0.4 170 66%128 (8, 8, 2) 654 0.2 68 82%
The method seems robust: our optimization choice is validated.
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 13 / 26
Physical background MINOS solver MINARET solver Conclusions
Numerical results with a PWR 900MW
Diffusion approximation, 2 energy groups, cell by cell, RTN0,289× 289× 60 unit meshes: 40× 106 unknowns,εf = 10−5.
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 14 / 26
Physical background MINOS solver MINARET solver Conclusions
Power maps
keff = 1.230157
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 15 / 26
Physical background MINOS solver MINARET solver Conclusions
GPU and DD (Titane, CCRT)
[Jamelot, Dubois, Lautard, Calvin, Baudron M&C 2011].
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 16 / 26
Physical background MINOS solver MINARET solver Conclusions
Kinetics diffusion equations
∀(x, t) ∈ R× [0, T ], ∀g ∈ 1, ..., G, ∀l ∈ 1, ...L :
Dggrad φg(x, t) + pg(x, t) = 0,
1
vg∂φg
∂t(x, t)− divpg(x, t) + σg
t (x)φg(x, t)
=G∑
g′=1
Sgg′
s (x)φg′
(x, t) +
G∑
g′=1
Sg′
f (x)φg′
(x, t) +
L∑
l=1
χgl (x)λlCl(x, t),
∂Cl
∂t(x, t) = −λlCl(x, t) +
G∑
g′=1
Fg′
l φg′
(x, t).
• Cl delayed neutron precursor concentrations.
• λl decay constant of precursor l.
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 17 / 26
Physical background MINOS solver MINARET solver Conclusions
Parareal algorithm (O. Mula-Hernandez)
∂ty +A(t; y) = 0, t ∈ [0, T ), y(0) = y0.
y(Tn) is approched by Y kn . Recurrence relation :
Y 0n+1 = G
Tn+1
Tn,
Y k+1n+1 = G
Tn+1
Tn(Y k+1
n ) + FTn+1
Tn(Y k
n )− GTn+1
Tn(Y k
n ),
0 10 20 30 40 500
2
4
6
8
Number of processors (N)
Spe
ed−
up
D, S−SMS, S−SD, W−SMS, W−S
More details in [Baudron, Lautard, Maday, Mula-Hernandez (DD21) 2012]
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 18 / 26
Physical background MINOS solver MINARET solver Conclusions
Description of MINARET (J.-J. Lautard, [Moller 2011])
• Criticality calculations, source problem and kinetics under progress.• Cylindrical mesh, unstructured in the (x, y) plane.• Energy E: multigroup theory.Eg ∈ [EG, EG−1] ∪ ... ∪ [E1, E0], EG < E0.
• Angular discretization Ω :SN transport (discrete ordinates method).SPN eqs and diffusion.
• Space discretization x : DGFEM.
• C++, librairie boost, MPI.
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 19 / 26
Physical background MINOS solver MINARET solver Conclusions
SN neutron transport equations
Ωd ∈ S2, d ∈ 1, ..., D. Solve in ψgd(x) such that:
Ωd .gradψgd(x)+ σg
t (x)ψgd(x) =
G∑
g′=0
∞∑
l=0
σg′→gs,l (x)
l∑
m=−l
Y ml (Ωd)φ
g′
lm(x)
+1
keff
χg
4π
G∑
g′=0
νg′
σg′
f (x)φg′
(x),
φgl,m(x) =
D∑
d=1
Y ml (Ωd)ψ
gd(x), φg(x) =
D∑
d=1
ωd ψgd(x).
Source iterationsAngular iterations d = 1 à D⇒ Jacobi
Spatial sweeping:
Ωd .gradψgd(x)+ σt(x)ψ
gd(x) = S(x, ψg′
d′ ,Ωd).
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 20 / 26
Physical background MINOS solver MINARET solver Conclusions
SN neutron transport equations
Ωd ∈ S2, d ∈ 1, ..., D. Solve in ψgd(x) such that:
Ωd .gradψgd(x)+ σg
t (x)ψgd(x) =
G∑
g′=0
∞∑
l=0
σg′→gs,l (x)
l∑
m=−l
Y ml (Ωd)φ
g′
lm(x)
+1
keff
χg
4π
G∑
g′=0
νg′
σg′
f (x)φg′
(x),
φgl,m(x) =
D∑
d=1
Y ml (Ωd)ψ
gd(x), φg(x) =
D∑
d=1
ωd ψgd(x).
Source iterationsAngular iterations d = 1 à D⇒ Distributed computing
Spatial sweeping:
Ωd .gradψgd(x)+ σt(x)ψ
gd(x) = S(x, ψg′
d′ ,Ωd).
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 20 / 26
Physical background MINOS solver MINARET solver Conclusions
MINARET solver algorithm (criticality)
SN solverEnergy sweeping
Outer iterations
of keffComputationSource
computation fission sourceUpdate of the
Acceleration of the outer iterations by means of Chebychev polynomials.DSA to accelerate the source iteration [Moller, Lautard, M&C 2011].
Parallelism: distributed computing.
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 21 / 26
Physical background MINOS solver MINARET solver Conclusions
J. Horowitz experimental core reactor
6 energy groups, 2D, 48× 103 cells, DG P1.
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 22 / 26
Physical background MINOS solver MINARET solver Conclusions
Numerical results.
6 energy groups, 2D, 48× 103 cells, DG P1, εf = 10−4.Opteron, 30 nodes, 4 proc. by nodes, 4 cores by proc., 2.6 Ghz.gcc 4.1.1 compiler, MPI Voltaire/Infiny Band 1.2.6 version.
keff = 1.31305S4, D = 12, 10.4M unknownsN Nd CPU (s) Eff.1 12 219 100%2 6 142 77%4 3 96 56%8 2 79 34%12 1 68 27%
keff = 1.31314S8, D = 40, 34.6M unknownsN Nd CPU (s) Eff.1 40 647 100%2 20 371 87%4 10 213 76%8 5 127 64%12 4 116 47%16 3 97 40%20 2 89 35%40 1 85 18%
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 23 / 26
Physical background MINOS solver MINARET solver Conclusions
CCRT Challenge: ESFR core reactor
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 24 / 26
Physical background MINOS solver MINARET solver Conclusions
Conclusions and perspectives
Numerical analysis:
• MINOS: DD for the source problem and kinetics (under work).
• MINOS kinetics: coupling DD and parareal [O. Mula-Hernandez].
• MINOS: Non conforming grids.
• MINARET kinetics: parareal [O. Mula-Hernandez].
• MINARET: unstructured 3D mesh.
• Coupling the solvers.
Computer engineering:
• MINOS: DD on hexagonal meshes.
• Data parallelism.
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 25 / 26
Physical background MINOS solver MINARET solver Conclusions
Thank you!
CEA/DEN/DANS/DM2S/SERMA/LLPR – E. Jamelot – HPC for neutronics – 20/12/12 – 26 / 26