Transcript Parallel Computing

Parallel Computing
Michael Young, Mark Iredell
NWS Computer History

1968 CDC 6600

1974 IBM 360

1983 CYBER 205
first vector parallelism

1991 Cray Y-MP
first shared memory parallelism

1994 Cray C-90
~16 gigaflops

2000 IBM SP
first distributed memory parallelism

2002 IBM SP P3

2004 IBM SP P4

2006 IBM SP P5

2009 IBM SP P6

2013 IBM Idataplex
SB
~200 teraflops
NEMS/GFS Modeling Summer School
2
Algorithm of the GFS Spectral Model

One time loop is divided into :
 Computation of the tendencies of
divergence, surface pressure, temperature
and vorticity and tracers (grid)
 Semi-implicit time integration (spectral)
 First half of time filter (spectral)
 Physical effects included in the model (grid)
 Damping to simulate subgrid dissipation
(spectral)
 Completion of the time filter (spectral)
NEMS/GFS Modeling Summer School
3
Algorithm of the GFS Spectral Model
Definitions :
Operational Spectral Truncation T574 with a
Physical Grid of 1760 longitudes by 880 latitudes and
64 vertical levels (23 km resolution)
θ is latitude
λ is longitude
l is zonal wavenumber
n is total wavenumber (zonal + meridional)
NEMS/GFS Modeling Summer School
4
Three Variable Spaces
Spectral (L x N x K)
 Fourier (L x J x K)
 Physical Grid ( I x J x k)
I is number of longitude points
J is number of latitudes
K is number of levels

NEMS/GFS Modeling Summer School
5
The Spectral Technique
All fields possess a spherical harmonic
representation:


JJ

l l
il

F
(, )

f
P
(sin
)
e


nn
l

0
n
l
where
1
2
l

n 2
n


1
(
2
n

1
)(
n

l
)
d
(
1

x
)
l
2
l
2
P
(
x
)

(
1

x
)
n


n
l

n
2
n
!
2
(
n

l
)!
dx


NEMS/GFS Modeling Summer School
6
Spectral to Grid Transform
Legendre transform:
J
F
(

)
fP
(sin

)

l
n

l
l l
nn
Fourier transform using FFT:
J
F
(

,
)
F
(

)
e

l
il

l
0
NEMS/GFS Modeling Summer School
7
Grid to Spectral Transform
1
l
fn 
2

2
0

l
il
F
(

,

)
P
(sin

)
e
cos dd
n

2
2
Inverse Fourier transform (FFT):
1 2
F()  F
(
,)eild

0
2

l
M
jl


i
M

F
(
)e
j,
j
0
Inverse Legendre (Gaussian quadrature):
l
l
l
F

w
F
(

)
P
(sin

)
n 
k
k n
k
N
k

1
NEMS/GFS Modeling Summer School
8
MPI and OpenMP



GFS uses Hybrid 1-Dimensional MPI layout
and OpenMP threading at do loop level
MPI (Message Passing Interface) is used to
communicate between tasks which contain a
subgrid of a field
OpenMP supports shared memory
multiprocessor programming (threading) using
compiler directives
NEMS/GFS Modeling Summer School
9
MPI and OpenMP


Data Transposes are implemented using
MPI_alltoallv
Required to switch between the variable
spaces which have different 1-D MPI
decompositions
NEMS/GFS Modeling Summer School
10
Spectral to Physical Grid



Call sumfln_slg_gg (Legendre Transform)
Call four_to_grid (FFT)
Data Transpose after Legendre Transform in
preparation for FFT to Physical grid space
call mpi_alltoallv(works,sendcounts,sdispls,mpi_r_mpi,
x
workr,recvcounts,sdispls,mpi_r_mpi,
x
mc_comp,ierr)
NEMS/GFS Modeling Summer School
11
Physical Grid to Spectral



Call Grid_to_four (Inverse FFT)
Call Four2fln_gg (Inverse Legendre Transform)
Data Transpose performed before the Inverse
Legendre Transform
call mpi_alltoallv(works,sendcounts,sdispls,MPI_R_MPI,
x
workr,recvcounts,sdispls,MPI_R_MPI,
x
MC_COMP,ierr)
NEMS/GFS Modeling Summer School
12
Physical Grid Space Parallelism

1-D MPI distributed over latitudes.
OpenMP threading used on longitude
points.
 Each MPI task holds a group of latitudes,
all longitudes, and all levels
 Cyclic distribution of latitudes used for
load balancing the MPI tasks due to a
smaller number of longitude points per
latitude as latitude increases
(approaches the poles).
NEMS/GFS Modeling Summer School
13
Physical Grid Space Parallelism

Cyclic distribution of latitudes example
5 MPI tasks and 20 Latitudes would be
Task 1 2 3 4 5
Lat 1 2 3 4 5
Lat 10 9 8 7 6
Lat 11 12 13 14 15
Lat 20 19 18 17 16
NEMS/GFS Modeling Summer School
14
Physical Grid Space Parallelism
Physical Grid Vector Length per OpenMP thread


NGPTC (namelist variable) defines number (block)
of longitude points per group (vector length per
processor) that each thread will work on
Typically set anywhere from 15-30 points
NEMS/GFS Modeling Summer School
15
Spectral Space Parallelism

Hybrid 1-D MPI layout with OpenMP threading
 Spectral space 1-D MPI distributed over
zonal wave numbers (l's). OpenMP
threading used on a stack of variables
times number of levels.
 Each MPI task holds a group of l’s, all n’s,
and all levels

Cyclic distribution of l's used for load balancing
the MPI tasks due to smaller numbers of
meridional points per zonal wave number as the
wave number increases.
NEMS/GFS Modeling Summer School
16
GFS Scalability



1-D MPI scales to 2/3 of the spectral
truncation. For T574 about 400 MPI tasks.
OpenMP threading scales to 8 threads.
T574 scales to 400 x 8 = 3200 processors.
NEMS/GFS Modeling Summer School
17