1 | <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN"> |
---|
2 | <html> |
---|
3 | <head> |
---|
4 | <meta content="text/html; charset=windows-1252" |
---|
5 | http-equiv="CONTENT-TYPE"> |
---|
6 | <title>PALM chapter 2.0</title> |
---|
7 | <meta content="StarOffice 7 (Win32)" name="GENERATOR"> |
---|
8 | <meta content="Marcus Oliver Letzel" name="AUTHOR"> |
---|
9 | <meta content="20040719;14534028" name="CREATED"> |
---|
10 | <meta content="20041117;10385730" name="CHANGED"> |
---|
11 | <meta content="parallel LES model" name="KEYWORDS"> |
---|
12 | <style> |
---|
13 | <!-- |
---|
14 | @page { size: 21cm 29.7cm } |
---|
15 | --> |
---|
16 | </style> |
---|
17 | </head> |
---|
18 | <body lang="en-US" dir="ltr"> |
---|
19 | <h2 style="line-height: 100%;"><font size="4">2.0 Basic techniques of |
---|
20 | the LES model and its parallelization </font> |
---|
21 | </h2> |
---|
22 | <p style="line-height: 100%;">LES models generally permit the |
---|
23 | simulation of turbulent flows, whereby those eddies, that carry the |
---|
24 | main energy are resolved by the numerical grid. Only the |
---|
25 | effect of such turbulence elements with diameter equal to or smaller |
---|
26 | than the grid spacing are parameterized in the model and |
---|
27 | by so-called subgrid-scale (SGS) transport. Larger structures are |
---|
28 | simulated directly (they are explicitly resolved) and their effects are |
---|
29 | represented by the advection terms. </p> |
---|
30 | <p style="font-style: normal; line-height: 100%;">PALM is based on the |
---|
31 | non-hydrostatic incompressible Boussinesq equations. It contains a |
---|
32 | water cycle with cloud formation and takes into account infrared |
---|
33 | radiative cooling in cloudy conditions. The model has six prognostic |
---|
34 | quantities in total u,v,w, liquid water potential temperature |
---|
35 | <font face="Thorndale, serif">Θ</font><sub>l </sub>(BETTS, |
---|
36 | 1973), total water content q and subgrid-scale turbulent kinetic energy |
---|
37 | e. The |
---|
38 | subgrid-scale turbulence is modeled according to DEARDOFF (1980) and |
---|
39 | requires the solution of an additional prognostic equation for the |
---|
40 | turbulent kinetic energy e. The long wave radiation scheme is based |
---|
41 | on the parametrization of cloud effective emissivity (e.g. Cox, 1976) |
---|
42 | and condensation is considered by a simple '0%-or-100%'-scheme, which |
---|
43 | assumes that within each grid box the air is either entirely |
---|
44 | unsaturated or entirely saturated ( see e.g., CUIJPERS and DUYNKERKE, |
---|
45 | 1993). The water cycle is closed by using a simplified version of |
---|
46 | KESSLERs scheme (KESSLER, 1965; 1969) to parameterize precipitation |
---|
47 | processes (MÜLLER and CHLOND, 1996). Incompressibility is |
---|
48 | applied by means of a Poisson equation for pressure, which is solved |
---|
49 | with a direct method (SCHUMANN and SWEET, 1988). The Poisson equation |
---|
50 | is Fourier transformed in both horizontal directions and the |
---|
51 | resulting tridiagonal matrix is solved for the transformed pressure |
---|
52 | which is then transformed back. Alternatively, a multigrid method can |
---|
53 | also be used. Lateral boundary conditions of the model are cyclic and |
---|
54 | MONIN-OBUKHOV similarity is assumed between the surface and the first |
---|
55 | computational grid level above. Alternatively, noncyclic boundary |
---|
56 | conditions |
---|
57 | (Dirichlet/Neumann) can be used along one of the |
---|
58 | horizontal directions. At the lower surface, either temperature/ |
---|
59 | humidity or their respective fluxes can be prescribed. </p> |
---|
60 | <p style="font-style: normal; line-height: 100%;">The advection terms |
---|
61 | are treated by the scheme proposed by PIACSEK and WILLIAMS (1970), |
---|
62 | which conserves the integral of linear and quadratic quantities up to |
---|
63 | very small errors. The advection of scalar quantities can optionally |
---|
64 | be performed by the monotone, locally modified version of Botts |
---|
65 | advection scheme (CHLOND, 1994). The time integration is performed |
---|
66 | with the third-order Runge-Kutta scheme. A second-order Runge-Kutta |
---|
67 | scheme, a leapfrog scheme and an Euler scheme are also implemented.</p> |
---|
68 | <p style="line-height: 100%;">By default, the time step is computed |
---|
69 | with respect to the different criteria (CFL, diffusion) and adapted |
---|
70 | automatically. In case of a non-zero geostrophic |
---|
71 | wind the coordinate system can be moved along with the mean wind in |
---|
72 | order to maximize the time step (Galilei-Transformation). </p> |
---|
73 | <p style="font-style: normal; line-height: 100%;">In principle a model |
---|
74 | run is carried out in the following way: After reading the control |
---|
75 | parameters given by the user, all prognostic variables are |
---|
76 | initialized. Initial values can be e.g. vertical profiles of the |
---|
77 | horizontal wind, calculated using a 1D subset of the 3D prognostic |
---|
78 | equation and are set in the 3D-Model as horizontally homogeneous |
---|
79 | initial values. Temperature profiles can only be prescribed linear |
---|
80 | (with constant gradients, which may change for different vertical |
---|
81 | height intervals) and they are assumed in the 1D-Model as stationary. |
---|
82 | After the initialization phase during which also different kinds of |
---|
83 | disturbances may be imposed to the prognostic fields, the time |
---|
84 | integration begins. Here for each individual time step the prognostic |
---|
85 | equations are successively solved for the velocity components u, v and |
---|
86 | w |
---|
87 | as well as for the potential temperature and possibly for the TKE. |
---|
88 | After the calculation of the boundary values in accordance with the |
---|
89 | given boundary conditions the provisional velocity fields are |
---|
90 | corrected with the help of the pressure solver. Following this, all |
---|
91 | diagnostic turbulence quantities including possible |
---|
92 | Prandtl-layerquantities are computed. At the end of a time |
---|
93 | step the data output requested by the user is made |
---|
94 | (e.g. statistic of analyses for control purposes or profiles and/or |
---|
95 | graphics data). If the given end-time was reached, binary data maybe |
---|
96 | be saved for restart. </p> |
---|
97 | <p style="font-style: normal; line-height: 100%;">The model is based |
---|
98 | on the originally non-parallel LES model which has been operated at the |
---|
99 | institute since 1989 |
---|
100 | and which was parallelized for massively parallel computers with |
---|
101 | distributed memory using the Message-Passing-Standard MPI. It is |
---|
102 | still applicable on a single processor and also well optimized for |
---|
103 | vector machines. The parallelization takes place via a so-called domain |
---|
104 | decomposition, which divides the entire model |
---|
105 | domain into individual, vertically standing cubes, which extend from |
---|
106 | the bottom to the top of the model domain. One processor (processing |
---|
107 | element, PE) is assigned to each cube, which |
---|
108 | accomplishes the computations on all grid points of the subdomain. |
---|
109 | Users can choose between a two- and a one-dimensional domain |
---|
110 | decomposition. A 1D-decomposition is preferred on machines with a |
---|
111 | slow network interconnection. In case of a 1D-decomposition, the |
---|
112 | grid points along x direction are |
---|
113 | distributed among the individual processors, but in y- and z-direction |
---|
114 | all respective grid points belong to the same PE. </p> |
---|
115 | <p style="line-height: 100%;">The calculation of central differences or |
---|
116 | non-local arithmetic operations (e.g. global |
---|
117 | sums, FFT) demands communication and an appropriate data exchange |
---|
118 | between the PEs. As a substantial innovation in relation to |
---|
119 | the non-parallel model version the individual subdomains are |
---|
120 | surrounded by so-called ghost points, which contain the grid point |
---|
121 | information of the neighbor processors. The appropriate grid point |
---|
122 | values must be exchanged after each change (i.e. in particular after |
---|
123 | each time step). For this purpose MPI routines (<tt>MPI_SENDRCV</tt>) |
---|
124 | are used. For the solution of the FFT conventional (non-parallelized) |
---|
125 | procedures are used. Given that the FFTs are used in x and/or |
---|
126 | y-direction, the data which lie distributed on the individual central |
---|
127 | processing elements, have to be collected and/or relocated before. |
---|
128 | This happens by means of the routine <tt>MPI_ALLTOALLV</tt>. Certain |
---|
129 | global operations like e.g. the search for absolute maxima or minima |
---|
130 | within the 3D-arrays likewise require the employment of special MPI |
---|
131 | routines (<tt>MPI_ALLREDUCE</tt>). </p> |
---|
132 | <p style="line-height: 100%;">Further details of the internal model |
---|
133 | structure are described in the <a href="../tec/index.html">technical/numerical |
---|
134 | documentation</a>. <br> |
---|
135 | </p> |
---|
136 | <hr><font color="#000080"><font color="#000080"><br> |
---|
137 | <a href="chapter_1.0.html"><font color="#000080"><img width="32" |
---|
138 | height="32" border="2" align="bottom" name="Grafik1" src="left.gif"></font></a><a |
---|
139 | href="index.html"><font color="#000080"><img width="32" height="32" |
---|
140 | border="2" align="bottom" name="Grafik2" src="up.gif"></font></a><a |
---|
141 | href="chapter_3.0.html"><font color="#000080"><img width="32" |
---|
142 | height="32" border="2" align="bottom" name="Grafik3" src="right.gif"></font></a><br> |
---|
143 | </font></font><br> |
---|
144 | <p style="line-height: 100%;"><span style="font-style: italic;">Last |
---|
145 | change: </span>14/04/05 (SR)<font color="#000080"><font color="#000080"><br> |
---|
146 | </font></font></p> |
---|
147 | </body> |
---|
148 | </html> |
---|