/[MITgcm]/manual/s_software/text/sarch.tex
ViewVC logotype

Contents of /manual/s_software/text/sarch.tex

Parent Directory Parent Directory | Revision Log Revision Log | View Revision Graph Revision Graph


Revision 1.14 - (show annotations) (download) (as text)
Wed Jan 28 20:38:28 2004 UTC (21 years, 5 months ago) by afe
Branch: MAIN
Changes since 1.13: +9 -2 lines
File MIME type: application/x-tex
'nuther change

1 % $Header: /u/u3/gcmpack/manual/part4/sarch.tex,v 1.12 2004/01/28 19:59:03 afe Exp $
2
3 This chapter focuses on describing the {\bf WRAPPER} environment within which
4 both the core numerics and the pluggable packages operate. The description
5 presented here is intended to be a detailed exposition and contains significant
6 background material, as well as advanced details on working with the WRAPPER.
7 The tutorial sections of this manual (see sections
8 \ref{sect:tutorials} and \ref{sect:tutorialIII})
9 contain more succinct, step-by-step instructions on running basic numerical
10 experiments, of varous types, both sequentially and in parallel. For many
11 projects simply starting from an example code and adapting it to suit a
12 particular situation
13 will be all that is required.
14 The first part of this chapter discusses the MITgcm architecture at an
15 abstract level. In the second part of the chapter we described practical
16 details of the MITgcm implementation and of current tools and operating system
17 features that are employed.
18
19 \section{Overall architectural goals}
20
21 Broadly, the goals of the software architecture employed in MITgcm are
22 three-fold
23
24 \begin{itemize}
25 \item We wish to be able to study a very broad range
26 of interesting and challenging rotating fluids problems.
27 \item We wish the model code to be readily targeted to
28 a wide range of platforms
29 \item On any given platform we would like to be
30 able to achieve performance comparable to an implementation
31 developed and specialized specifically for that platform.
32 \end{itemize}
33
34 These points are summarized in figure \ref{fig:mitgcm_architecture_goals}
35 which conveys the goals of the MITgcm design. The goals lead to
36 a software architecture which at the high-level can be viewed as consisting
37 of
38
39 \begin{enumerate}
40 \item A core set of numerical and support code. This is discussed in detail in
41 section \ref{sect:partII}.
42 \item A scheme for supporting optional "pluggable" {\bf packages} (containing
43 for example mixed-layer schemes, biogeochemical schemes, atmospheric physics).
44 These packages are used both to overlay alternate dynamics and to introduce
45 specialized physical content onto the core numerical code. An overview of
46 the {\bf package} scheme is given at the start of part \ref{part:packages}.
47 \item A support framework called {\bf WRAPPER} (Wrappable Application Parallel
48 Programming Environment Resource), within which the core numerics and pluggable
49 packages operate.
50 \end{enumerate}
51
52 This chapter focuses on describing the {\bf WRAPPER} environment under which
53 both the core numerics and the pluggable packages function. The description
54 presented here is intended to be a detailed exposition and contains significant
55 background material, as well as advanced details on working with the WRAPPER.
56 The examples section of this manual (part \ref{part:example}) contains more
57 succinct, step-by-step instructions on running basic numerical
58 experiments both sequentially and in parallel. For many projects simply
59 starting from an example code and adapting it to suit a particular situation
60 will be all that is required.
61
62
63 \begin{figure}
64 \begin{center}
65 \resizebox{!}{2.5in}{\includegraphics{part4/mitgcm_goals.eps}}
66 \end{center}
67 \caption{
68 The MITgcm architecture is designed to allow simulation of a wide
69 range of physical problems on a wide range of hardware. The computational
70 resource requirements of the applications targeted range from around
71 $10^7$ bytes ( $\approx 10$ megabytes ) of memory to $10^{11}$ bytes
72 ( $\approx 100$ gigabytes). Arithmetic operation counts for the applications of
73 interest range from $10^{9}$ floating point operations to more than $10^{17}$
74 floating point operations.}
75 \label{fig:mitgcm_architecture_goals}
76 \end{figure}
77
78 \section{WRAPPER}
79
80 A significant element of the software architecture utilized in
81 MITgcm is a software superstructure and substructure collectively
82 called the WRAPPER (Wrappable Application Parallel Programming
83 Environment Resource). All numerical and support code in MITgcm is written
84 to ``fit'' within the WRAPPER infrastructure. Writing code to ``fit'' within
85 the WRAPPER means that coding has to follow certain, relatively
86 straightforward, rules and conventions ( these are discussed further in
87 section \ref{sect:specifying_a_decomposition} ).
88
89 The approach taken by the WRAPPER is illustrated in figure
90 \ref{fig:fit_in_wrapper} which shows how the WRAPPER serves to insulate code
91 that fits within it from architectural differences between hardware platforms
92 and operating systems. This allows numerical code to be easily retargetted.
93
94
95 \begin{figure}
96 \begin{center}
97 \resizebox{!}{4.5in}{\includegraphics{part4/fit_in_wrapper.eps}}
98 \end{center}
99 \caption{
100 Numerical code is written to fit within a software support
101 infrastructure called WRAPPER. The WRAPPER is portable and
102 can be specialized for a wide range of specific target hardware and
103 programming environments, without impacting numerical code that fits
104 within the WRAPPER. Codes that fit within the WRAPPER can generally be
105 made to run as fast on a particular platform as codes specially
106 optimized for that platform.}
107 \label{fig:fit_in_wrapper}
108 \end{figure}
109
110 \subsection{Target hardware}
111 \label{sect:target_hardware}
112
113 The WRAPPER is designed to target as broad as possible a range of computer
114 systems. The original development of the WRAPPER took place on a
115 multi-processor, CRAY Y-MP system. On that system, numerical code performance
116 and scaling under the WRAPPER was in excess of that of an implementation that
117 was tightly bound to the CRAY systems proprietary multi-tasking and
118 micro-tasking approach. Later developments have been carried out on
119 uniprocessor and multi-processor Sun systems with both uniform memory access
120 (UMA) and non-uniform memory access (NUMA) designs. Significant work has also
121 been undertaken on x86 cluster systems, Alpha processor based clustered SMP
122 systems, and on cache-coherent NUMA (CC-NUMA) systems from Silicon Graphics.
123 The MITgcm code, operating within the WRAPPER, is also routinely used on
124 large scale MPP systems (for example T3E systems and IBM SP systems). In all
125 cases numerical code, operating within the WRAPPER, performs and scales very
126 competitively with equivalent numerical code that has been modified to contain
127 native optimizations for a particular system \ref{ref hoe and hill, ecmwf}.
128
129 \subsection{Supporting hardware neutrality}
130
131 The different systems listed in section \ref{sect:target_hardware} can be
132 categorized in many different ways. For example, one common distinction is
133 between shared-memory parallel systems (SMP's, PVP's) and distributed memory
134 parallel systems (for example x86 clusters and large MPP systems). This is one
135 example of a difference between compute platforms that can impact an
136 application. Another common distinction is between vector processing systems
137 with highly specialized CPU's and memory subsystems and commodity
138 microprocessor based systems. There are numerous other differences, especially
139 in relation to how parallel execution is supported. To capture the essential
140 differences between different platforms the WRAPPER uses a {\it machine model}.
141
142 \subsection{WRAPPER machine model}
143
144 Applications using the WRAPPER are not written to target just one
145 particular machine (for example an IBM SP2) or just one particular family or
146 class of machines (for example Parallel Vector Processor Systems). Instead the
147 WRAPPER provides applications with an
148 abstract {\it machine model}. The machine model is very general, however, it can
149 easily be specialized to fit, in a computationally efficient manner, any
150 computer architecture currently available to the scientific computing community.
151
152 \subsection{Machine model parallelism}
153
154 Codes operating under the WRAPPER target an abstract machine that is assumed to
155 consist of one or more logical processors that can compute concurrently.
156 Computational work is divided among the logical
157 processors by allocating ``ownership'' to
158 each processor of a certain set (or sets) of calculations. Each set of
159 calculations owned by a particular processor is associated with a specific
160 region of the physical space that is being simulated, only one processor will
161 be associated with each such region (domain decomposition).
162
163 In a strict sense the logical processors over which work is divided do not need
164 to correspond to physical processors. It is perfectly possible to execute a
165 configuration decomposed for multiple logical processors on a single physical
166 processor. This helps ensure that numerical code that is written to fit
167 within the WRAPPER will parallelize with no additional effort and is
168 also useful when debugging codes. Generally, however,
169 the computational domain will be subdivided over multiple logical
170 processors in order to then bind those logical processors to physical
171 processor resources that can compute in parallel.
172
173 \subsubsection{Tiles}
174
175 Computationally, associated with each region of physical
176 space allocated to a particular logical processor, there will be data
177 structures (arrays, scalar variables etc...) that hold the simulated state of
178 that region. We refer to these data structures as being {\bf owned} by the
179 processor to which their
180 associated region of physical space has been allocated. Individual
181 regions that are allocated to processors are called {\bf tiles}. A
182 processor can own more
183 than one tile. Figure \ref{fig:domaindecomp} shows a physical domain being
184 mapped to a set of logical processors, with each processors owning a single
185 region of the domain (a single tile). Except for periods of
186 communication and coordination, each processor computes autonomously, working
187 only with data from the tile (or tiles) that the processor owns. When multiple
188 tiles are alloted to a single processor, each tile is computed on
189 independently of the other tiles, in a sequential fashion.
190
191 \begin{figure}
192 \begin{center}
193 \resizebox{5in}{!}{
194 \includegraphics{part4/domain_decomp.eps}
195 }
196 \end{center}
197 \caption{ The WRAPPER provides support for one and two dimensional
198 decompositions of grid-point domains. The figure shows a hypothetical domain of
199 total size $N_{x}N_{y}N_{z}$. This hypothetical domain is decomposed in
200 two-dimensions along the $N_{x}$ and $N_{y}$ directions. The resulting {\bf
201 tiles} are {\bf owned} by different processors. The {\bf owning}
202 processors perform the
203 arithmetic operations associated with a {\bf tile}. Although not illustrated
204 here, a single processor can {\bf own} several {\bf tiles}.
205 Whenever a processor wishes to transfer data between tiles or
206 communicate with other processors it calls a WRAPPER supplied
207 function.
208 } \label{fig:domaindecomp}
209 \end{figure}
210
211
212 \subsubsection{Tile layout}
213
214 Tiles consist of an interior region and an overlap region. The overlap region
215 of a tile corresponds to the interior region of an adjacent tile.
216 In figure \ref{fig:tiledworld} each tile would own the region
217 within the black square and hold duplicate information for overlap
218 regions extending into the tiles to the north, south, east and west.
219 During
220 computational phases a processor will reference data in an overlap region
221 whenever it requires values that outside the domain it owns. Periodically
222 processors will make calls to WRAPPER functions to communicate data between
223 tiles, in order to keep the overlap regions up to date (see section
224 \ref{sect:communication_primitives}). The WRAPPER functions can use a
225 variety of different mechanisms to communicate data between tiles.
226
227 \begin{figure}
228 \begin{center}
229 \resizebox{5in}{!}{
230 \includegraphics{part4/tiled-world.eps}
231 }
232 \end{center}
233 \caption{ A global grid subdivided into tiles.
234 Tiles contain a interior region and an overlap region.
235 Overlap regions are periodically updated from neighboring tiles.
236 } \label{fig:tiledworld}
237 \end{figure}
238
239 \subsection{Communication mechanisms}
240
241 Logical processors are assumed to be able to exchange information
242 between tiles and between each other using at least one of two possible
243 mechanisms.
244
245 \begin{itemize}
246 \item {\bf Shared memory communication}.
247 Under this mode of communication data transfers are assumed to be possible
248 using direct addressing of regions of memory. In this case a CPU is able to read
249 (and write) directly to regions of memory "owned" by another CPU
250 using simple programming language level assignment operations of the
251 the sort shown in figure \ref{fig:simple_assign}. In this way one CPU
252 (CPU1 in the figure) can communicate information to another CPU (CPU2 in the
253 figure) by assigning a particular value to a particular memory location.
254
255 \item {\bf Distributed memory communication}.
256 Under this mode of communication there is no mechanism, at the application code level,
257 for directly addressing regions of memory owned and visible to another CPU. Instead
258 a communication library must be used as illustrated in figure
259 \ref{fig:comm_msg}. In this case CPU's must call a function in the API of the
260 communication library to communicate data from a tile that it owns to a tile
261 that another CPU owns. By default the WRAPPER binds to the MPI communication
262 library \ref{MPI} for this style of communication.
263 \end{itemize}
264
265 The WRAPPER assumes that communication will use one of these two styles
266 of communication. The underlying hardware and operating system support
267 for the style used is not specified and can vary from system to system.
268
269 \begin{figure}
270 \begin{verbatim}
271
272 CPU1 | CPU2
273 ==== | ====
274 |
275 a(3) = 8 | WHILE ( a(3) .NE. 8 )
276 | WAIT
277 | END WHILE
278 |
279 \end{verbatim}
280 \caption{ In the WRAPPER shared memory communication model, simple writes to an
281 array can be made to be visible to other CPU's at the application code level.
282 So that for example, if one CPU (CPU1 in the figure above) writes the value $8$ to
283 element $3$ of array $a$, then other CPU's (for example CPU2 in the figure above)
284 will be able to see the value $8$ when they read from $a(3)$.
285 This provides a very low latency and high bandwidth communication
286 mechanism.
287 } \label{fig:simple_assign}
288 \end{figure}
289
290 \begin{figure}
291 \begin{verbatim}
292
293 CPU1 | CPU2
294 ==== | ====
295 |
296 a(3) = 8 | WHILE ( a(3) .NE. 8 )
297 CALL SEND( CPU2,a(3) ) | CALL RECV( CPU1, a(3) )
298 | END WHILE
299 |
300 \end{verbatim}
301 \caption{ In the WRAPPER distributed memory communication model
302 data can not be made directly visible to other CPU's.
303 If one CPU writes the value $8$ to element $3$ of array $a$, then
304 at least one of CPU1 and/or CPU2 in the figure above will need
305 to call a bespoke communication library in order for the updated
306 value to be communicated between CPU's.
307 } \label{fig:comm_msg}
308 \end{figure}
309
310 \subsection{Shared memory communication}
311 \label{sect:shared_memory_communication}
312
313 Under shared communication independent CPU's are operating
314 on the exact same global address space at the application level.
315 This means that CPU 1 can directly write into global
316 data structures that CPU 2 ``owns'' using a simple
317 assignment at the application level.
318 This is the model of memory access is supported at the basic system
319 design level in ``shared-memory'' systems such as PVP systems, SMP systems,
320 and on distributed shared memory systems (the SGI Origin).
321 On such systems the WRAPPER will generally use simple read and write statements
322 to access directly application data structures when communicating between CPU's.
323
324 In a system where assignments statements, like the one in figure
325 \ref{fig:simple_assign} map directly to
326 hardware instructions that transport data between CPU and memory banks, this
327 can be a very efficient mechanism for communication. In this case two CPU's,
328 CPU1 and CPU2, can communicate simply be reading and writing to an
329 agreed location and following a few basic rules. The latency of this sort
330 of communication is generally not that much higher than the hardware
331 latency of other memory accesses on the system. The bandwidth available
332 between CPU's communicating in this way can be close to the bandwidth of
333 the systems main-memory interconnect. This can make this method of
334 communication very efficient provided it is used appropriately.
335
336 \subsubsection{Memory consistency}
337 \label{sect:memory_consistency}
338
339 When using shared memory communication between
340 multiple processors the WRAPPER level shields user applications from
341 certain counter-intuitive system behaviors. In particular, one issue the
342 WRAPPER layer must deal with is a systems memory model. In general the order
343 of reads and writes expressed by the textual order of an application code may
344 not be the ordering of instructions executed by the processor performing the
345 application. The processor performing the application instructions will always
346 operate so that, for the application instructions the processor is executing,
347 any reordering is not apparent. However, in general machines are often
348 designed so that reordering of instructions is not hidden from other second
349 processors. This means that, in general, even on a shared memory system two
350 processors can observe inconsistent memory values.
351
352 The issue of memory consistency between multiple processors is discussed at
353 length in many computer science papers, however, from a practical point of
354 view, in order to deal with this issue, shared memory machines all provide
355 some mechanism to enforce memory consistency when it is needed. The exact
356 mechanism employed will vary between systems. For communication using shared
357 memory, the WRAPPER provides a place to invoke the appropriate mechanism to
358 ensure memory consistency for a particular platform.
359
360 \subsubsection{Cache effects and false sharing}
361 \label{sect:cache_effects_and_false_sharing}
362
363 Shared-memory machines often have local to processor memory caches
364 which contain mirrored copies of main memory. Automatic cache-coherence
365 protocols are used to maintain consistency between caches on different
366 processors. These cache-coherence protocols typically enforce consistency
367 between regions of memory with large granularity (typically 128 or 256 byte
368 chunks). The coherency protocols employed can be expensive relative to other
369 memory accesses and so care is taken in the WRAPPER (by padding synchronization
370 structures appropriately) to avoid unnecessary coherence traffic.
371
372 \subsubsection{Operating system support for shared memory.}
373
374 Applications running under multiple threads within a single process can
375 use shared memory communication. In this case {\it all} the memory locations
376 in an application are potentially visible to all the compute threads. Multiple
377 threads operating within a single process is the standard mechanism for
378 supporting shared memory that the WRAPPER utilizes. Configuring and launching
379 code to run in multi-threaded mode on specific platforms is discussed in
380 section \ref{sect:running_with_threads}. However, on many systems, potentially
381 very efficient mechanisms for using shared memory communication between
382 multiple processes (in contrast to multiple threads within a single
383 process) also exist. In most cases this works by making a limited region of
384 memory shared between processes. The MMAP \ref{magicgarden} and
385 IPC \ref{magicgarden} facilities in UNIX systems provide this capability as do
386 vendor specific tools like LAPI \ref{IBMLAPI} and IMC \ref{Memorychannel}.
387 Extensions exist for the WRAPPER that allow these mechanisms
388 to be used for shared memory communication. However, these mechanisms are not
389 distributed with the default WRAPPER sources, because of their proprietary
390 nature.
391
392 \subsection{Distributed memory communication}
393 \label{sect:distributed_memory_communication}
394 Many parallel systems are not constructed in a way where it is
395 possible or practical for an application to use shared memory
396 for communication. For example cluster systems consist of individual computers
397 connected by a fast network. On such systems their is no notion of shared memory
398 at the system level. For this sort of system the WRAPPER provides support
399 for communication based on a bespoke communication library
400 (see figure \ref{fig:comm_msg}). The default communication library used is MPI
401 \ref{mpi}. However, it is relatively straightforward to implement bindings to
402 optimized platform specific communication libraries. For example the work
403 described in \ref{hoe-hill:99} substituted standard MPI communication for a
404 highly optimized library.
405
406 \subsection{Communication primitives}
407 \label{sect:communication_primitives}
408
409 \begin{figure}
410 \begin{center}
411 \resizebox{5in}{!}{
412 \includegraphics{part4/comm-primm.eps}
413 }
414 \end{center}
415 \caption{Three performance critical parallel primitives are provided
416 by the WRAPPER. These primitives are always used to communicate data
417 between tiles. The figure shows four tiles. The curved arrows indicate
418 exchange primitives which transfer data between the overlap regions at tile
419 edges and interior regions for nearest-neighbor tiles.
420 The straight arrows symbolize global sum operations which connect all tiles.
421 The global sum operation provides both a key arithmetic primitive and can
422 serve as a synchronization primitive. A third barrier primitive is also
423 provided, it behaves much like the global sum primitive.
424 } \label{fig:communication_primitives}
425 \end{figure}
426
427
428 Optimized communication support is assumed to be possibly available
429 for a small number of communication operations.
430 It is assumed that communication performance optimizations can
431 be achieved by optimizing a small number of communication primitives.
432 Three optimizable primitives are provided by the WRAPPER
433 \begin{itemize}
434 \item{\bf EXCHANGE} This operation is used to transfer data between interior
435 and overlap regions of neighboring tiles. A number of different forms of this
436 operation are supported. These different forms handle
437 \begin{itemize}
438 \item Data type differences. Sixty-four bit and thirty-two bit fields may be handled
439 separately.
440 \item Bindings to different communication methods.
441 Exchange primitives select between using shared memory or distributed
442 memory communication.
443 \item Transformation operations required when transporting
444 data between different grid regions. Transferring data
445 between faces of a cube-sphere grid, for example, involves a rotation
446 of vector components.
447 \item Forward and reverse mode computations. Derivative calculations require
448 tangent linear and adjoint forms of the exchange primitives.
449
450 \end{itemize}
451
452 \item{\bf GLOBAL SUM} The global sum operation is a central arithmetic
453 operation for the pressure inversion phase of the MITgcm algorithm.
454 For certain configurations scaling can be highly sensitive to
455 the performance of the global sum primitive. This operation is a collective
456 operation involving all tiles of the simulated domain. Different forms
457 of the global sum primitive exist for handling
458 \begin{itemize}
459 \item Data type differences. Sixty-four bit and thirty-two bit fields may be handled
460 separately.
461 \item Bindings to different communication methods.
462 Exchange primitives select between using shared memory or distributed
463 memory communication.
464 \item Forward and reverse mode computations. Derivative calculations require
465 tangent linear and adjoint forms of the exchange primitives.
466 \end{itemize}
467
468 \item{\bf BARRIER} The WRAPPER provides a global synchronization function
469 called barrier. This is used to synchronize computations over all tiles.
470 The {\bf BARRIER} and {\bf GLOBAL SUM} primitives have much in common and in
471 some cases use the same underlying code.
472 \end{itemize}
473
474
475 \subsection{Memory architecture}
476
477 The WRAPPER machine model is aimed to target efficiently systems with
478 highly pipelined memory architectures and systems with deep memory
479 hierarchies that favor memory reuse. This is achieved by supporting a
480 flexible tiling strategy as shown in figure \ref{fig:tiling-strategy}.
481 Within a CPU computations are carried out sequentially on each tile
482 in turn. By reshaping tiles according to the target platform it is
483 possible to automatically tune code to improve memory performance.
484 On a vector machine a given domain might be sub-divided into a few
485 long, thin regions. On a commodity microprocessor based system, however,
486 the same region could be simulated use many more smaller
487 sub-domains.
488
489
490 \begin{figure}
491 \begin{center}
492 \resizebox{5in}{!}{
493 \includegraphics{part4/tiling_detail.eps}
494 }
495 \end{center}
496 \caption{The tiling strategy that the WRAPPER supports allows tiles
497 to be shaped to suit the underlying system memory architecture.
498 Compact tiles that lead to greater memory reuse can be used on cache
499 based systems (upper half of figure) with deep memory hierarchies, long tiles
500 with large inner loops can be used to exploit vector systems having
501 highly pipelined memory systems.
502 } \label{fig:tiling-strategy}
503 \end{figure}
504
505 \newpage
506 \subsection{Summary}
507 Following the discussion above, the machine model that the WRAPPER
508 presents to an application has the following characteristics
509
510 \begin{itemize}
511 \item The machine consists of one or more logical processors. \vspace{-3mm}
512 \item Each processor operates on tiles that it owns.\vspace{-3mm}
513 \item A processor may own more than one tile.\vspace{-3mm}
514 \item Processors may compute concurrently.\vspace{-3mm}
515 \item Exchange of information between tiles is handled by the
516 machine (WRAPPER) not by the application.
517 \end{itemize}
518 Behind the scenes this allows the WRAPPER to adapt the machine model
519 functions to exploit hardware on which
520 \begin{itemize}
521 \item Processors may be able to communicate very efficiently with each other
522 using shared memory. \vspace{-3mm}
523 \item An alternative communication mechanism based on a relatively
524 simple inter-process communication API may be required.\vspace{-3mm}
525 \item Shared memory may not necessarily obey sequential consistency,
526 however some mechanism will exist for enforcing memory consistency.
527 \vspace{-3mm}
528 \item Memory consistency that is enforced at the hardware level
529 may be expensive. Unnecessary triggering of consistency protocols
530 should be avoided. \vspace{-3mm}
531 \item Memory access patterns may need to either repetitive or highly
532 pipelined for optimum hardware performance. \vspace{-3mm}
533 \end{itemize}
534
535 This generic model captures the essential hardware ingredients
536 of almost all successful scientific computer systems designed in the
537 last 50 years.
538
539 \section{Using the WRAPPER}
540
541 In order to support maximum portability the WRAPPER is implemented primarily
542 in sequential Fortran 77. At a practical level the key steps provided by the
543 WRAPPER are
544 \begin{enumerate}
545 \item specifying how a domain will be decomposed
546 \item starting a code in either sequential or parallel modes of operations
547 \item controlling communication between tiles and between concurrently
548 computing CPU's.
549 \end{enumerate}
550 This section describes the details of each of these operations.
551 Section \ref{sect:specifying_a_decomposition} explains how the way in which
552 a domain is decomposed (or composed) is expressed. Section
553 \ref{sect:starting_a_code} describes practical details of running codes
554 in various different parallel modes on contemporary computer systems.
555 Section \ref{sect:controlling_communication} explains the internal information
556 that the WRAPPER uses to control how information is communicated between
557 tiles.
558
559 \subsection{Specifying a domain decomposition}
560 \label{sect:specifying_a_decomposition}
561
562 At its heart much of the WRAPPER works only in terms of a collection of tiles
563 which are interconnected to each other. This is also true of application
564 code operating within the WRAPPER. Application code is written as a series
565 of compute operations, each of which operates on a single tile. If
566 application code needs to perform operations involving data
567 associated with another tile, it uses a WRAPPER function to obtain
568 that data.
569 The specification of how a global domain is constructed from tiles or alternatively
570 how a global domain is decomposed into tiles is made in the file {\em SIZE.h}.
571 This file defines the following parameters \\
572
573 \fbox{
574 \begin{minipage}{4.75in}
575 Parameters: {\em sNx, sNy, OLx, OLy, nSx, nSy, nPx, nPy} \\
576 File: {\em model/inc/SIZE.h}
577 \end{minipage}
578 } \\
579
580 Together these parameters define a tiling decomposition of the style shown in
581 figure \ref{fig:labelled_tile}. The parameters {\em sNx} and {\em sNy} define
582 the size of an individual tile. The parameters {\em OLx} and {\em OLy} define the
583 maximum size of the overlap extent. This must be set to the maximum width
584 of the computation stencil that the numerical code finite-difference operations
585 require between overlap region updates. The maximum overlap required
586 by any of the operations in the MITgcm code distributed with this release is three grid
587 points. This is set by the requirements of the $\nabla^4$ dissipation and
588 diffusion operator. Code modifications and enhancements that involve adding wide
589 finite-difference stencils may require increasing {\em OLx} and {\em OLy}.
590 Setting {\em OLx} and {\em OLy} to a too large value will decrease code
591 performance (because redundant computations will be performed), however it will
592 not cause any other problems.
593
594 \begin{figure}
595 \begin{center}
596 \resizebox{5in}{!}{
597 \includegraphics{part4/size_h.eps}
598 }
599 \end{center}
600 \caption{ The three level domain decomposition hierarchy employed by the
601 WRAPPER. A domain is composed of tiles. Multiple tiles can be allocated
602 to a single process. Multiple processes can exist, each with multiple tiles.
603 Tiles within a process can be spread over multiple compute threads.
604 } \label{fig:labelled_tile}
605 \end{figure}
606
607 The parameters {\em nSx} and {\em nSy} specify the number of tiles that will
608 be created within a single process. Each of these tiles will have internal
609 dimensions of {\em sNx} and {\em sNy}. If, when the code is executed, these tiles are
610 allocated to different threads of a process that are then bound to
611 different physical processors ( see the multi-threaded
612 execution discussion in section \ref{sect:starting_the_code} ) then
613 computation will be performed concurrently on each tile. However, it is also
614 possible to run the same decomposition within a process running a single thread on
615 a single processor. In this case the tiles will be computed over sequentially.
616 If the decomposition is run in a single process running multiple threads
617 but attached to a single physical processor, then, in general, the computation
618 for different tiles will be interleaved by system level software.
619 This too is a valid mode of operation.
620
621 The parameters {\em sNx, sNy, OLx, OLy, nSx} and{\em nSy} are used extensively by
622 numerical code. The settings of {\em sNx, sNy, OLx} and {\em OLy}
623 are used to form the loop ranges for many numerical calculations and
624 to provide dimensions for arrays holding numerical state.
625 The {\em nSx} and{\em nSy} are used in conjunction with the thread number
626 parameter {\em myThid}. Much of the numerical code operating within the
627 WRAPPER takes the form
628 \begin{verbatim}
629 DO bj=myByLo(myThid),myByHi(myThid)
630 DO bi=myBxLo(myThid),myBxHi(myThid)
631 :
632 a block of computations ranging
633 over 1,sNx +/- OLx and 1,sNy +/- OLy grid points
634 :
635 ENDDO
636 ENDDO
637
638 communication code to sum a number or maybe update
639 tile overlap regions
640
641 DO bj=myByLo(myThid),myByHi(myThid)
642 DO bi=myBxLo(myThid),myBxHi(myThid)
643 :
644 another block of computations ranging
645 over 1,sNx +/- OLx and 1,sNy +/- OLy grid points
646 :
647 ENDDO
648 ENDDO
649 \end{verbatim}
650 The variables {\em myBxLo(myThid), myBxHi(myThid), myByLo(myThid)} and {\em
651 myByHi(myThid)} set the bounds of the loops in {\em bi} and {\em bj} in this
652 schematic. These variables specify the subset of the tiles in
653 the range 1,{\em nSx} and 1,{\em nSy} that the logical processor bound to
654 thread number {\em myThid} owns. The thread number variable {\em myThid}
655 ranges from 1 to the total number of threads requested at execution time.
656 For each value of {\em myThid} the loop scheme above will step sequentially
657 through the tiles owned by that thread. However, different threads will
658 have different ranges of tiles assigned to them, so that separate threads can
659 compute iterations of the {\em bi}, {\em bj} loop concurrently.
660 Within a {\em bi}, {\em bj} loop
661 computation is performed concurrently over as many processes and threads
662 as there are physical processors available to compute.
663
664 An exception to the the use of {\em bi} and {\em bj} in loops arises in the
665 exchange routines used when the exch2 package is used with the cubed
666 sphere. In this case {\em bj} is generally set to 1 and the loop runs from
667 1,{\em bi}. Within the loop {\em bi} is used to retrieve the tile number,
668 which is then used to reference exchange parameters.
669
670 The amount of computation that can be embedded
671 a single loop over {\em bi} and {\em bj} varies for different parts of the
672 MITgcm algorithm. Figure \ref{fig:bibj_extract} shows a code extract
673 from the two-dimensional implicit elliptic solver. This portion of the
674 code computes the l2Norm of a vector whose elements are held in
675 the array {\em cg2d\_r} writing the final result to scalar variable
676 {\em err}. In this case, because the l2norm requires a global reduction,
677 the {\em bi},{\em bj} loop only contains one statement. This computation
678 phase is then followed by a communication phase in which all threads and
679 processes must participate. However,
680 in other areas of the MITgcm code entries subsections of code are within
681 a single {\em bi},{\em bj} loop. For example the evaluation of all
682 the momentum equation prognostic terms ( see {\em S/R DYNAMICS()})
683 is within a single {\em bi},{\em bj} loop.
684
685 \begin{figure}
686 \begin{verbatim}
687 REAL*8 cg2d_r(1-OLx:sNx+OLx,1-OLy:sNy+OLy,nSx,nSy)
688 REAL*8 err
689 :
690 :
691 other computations
692 :
693 :
694 err = 0.
695 DO bj=myByLo(myThid),myByHi(myThid)
696 DO bi=myBxLo(myThid),myBxHi(myThid)
697 DO J=1,sNy
698 DO I=1,sNx
699 err = err +
700 & cg2d_r(I,J,bi,bj)*cg2d_r(I,J,bi,bj)
701 ENDDO
702 ENDDO
703 ENDDO
704 ENDDO
705
706 CALL GLOBAL_SUM_R8( err , myThid )
707 err = SQRT(err)
708
709 \end{verbatim}
710 \caption{Example of numerical code for calculating
711 the l2-norm of a vector within the WRAPPER. Notice that
712 under the WRAPPER arrays such as {\em cg2d\_r} have two extra trailing
713 dimensions. These right most indices are tile indexes. Different
714 threads with a single process operate on different ranges of tile
715 index, as controlled by the settings of
716 {\em myByLo, myByHi, myBxLo} and {\em myBxHi}.
717 } \label{fig:bibj_extract}
718 \end{figure}
719
720 The final decomposition parameters are {\em nPx} and {\em nPy}. These parameters
721 are used to indicate to the WRAPPER level how many processes (each with
722 {\em nSx}$\times${\em nSy} tiles) will be used for this simulation.
723 This information is needed during initialization and during I/O phases.
724 However, unlike the variables {\em sNx, sNy, OLx, OLy, nSx} and {\em nSy}
725 the values of {\em nPx} and {\em nPy} are absent
726 from the core numerical and support code.
727
728 \subsubsection{Examples of {\em SIZE.h} specifications}
729
730 The following different {\em SIZE.h} parameter setting illustrate how to
731 interpret the values of {\em sNx, sNy, OLx, OLy, nSx, nSy, nPx}
732 and {\em nPy}.
733 \begin{enumerate}
734 \item
735 \begin{verbatim}
736 PARAMETER (
737 & sNx = 90,
738 & sNy = 40,
739 & OLx = 3,
740 & OLy = 3,
741 & nSx = 1,
742 & nSy = 1,
743 & nPx = 1,
744 & nPy = 1)
745 \end{verbatim}
746 This sets up a single tile with x-dimension of ninety grid points, y-dimension of
747 forty grid points, and x and y overlaps of three grid points each.
748 \item
749 \begin{verbatim}
750 PARAMETER (
751 & sNx = 45,
752 & sNy = 20,
753 & OLx = 3,
754 & OLy = 3,
755 & nSx = 1,
756 & nSy = 1,
757 & nPx = 2,
758 & nPy = 2)
759 \end{verbatim}
760 This sets up tiles with x-dimension of forty-five grid points, y-dimension of
761 twenty grid points, and x and y overlaps of three grid points each. There are
762 four tiles allocated to four separate processes ({\em nPx=2,nPy=2}) and
763 arranged so that the global domain size is again ninety grid points in x and
764 forty grid points in y. In general the formula for global grid size (held in
765 model variables {\em Nx} and {\em Ny}) is
766 \begin{verbatim}
767 Nx = sNx*nSx*nPx
768 Ny = sNy*nSy*nPy
769 \end{verbatim}
770 \item
771 \begin{verbatim}
772 PARAMETER (
773 & sNx = 90,
774 & sNy = 10,
775 & OLx = 3,
776 & OLy = 3,
777 & nSx = 1,
778 & nSy = 2,
779 & nPx = 1,
780 & nPy = 2)
781 \end{verbatim}
782 This sets up tiles with x-dimension of ninety grid points, y-dimension of
783 ten grid points, and x and y overlaps of three grid points each. There are
784 four tiles allocated to two separate processes ({\em nPy=2}) each of which
785 has two separate sub-domains {\em nSy=2},
786 The global domain size is again ninety grid points in x and
787 forty grid points in y. The two sub-domains in each process will be computed
788 sequentially if they are given to a single thread within a single process.
789 Alternatively if the code is invoked with multiple threads per process
790 the two domains in y may be computed concurrently.
791 \item
792 \begin{verbatim}
793 PARAMETER (
794 & sNx = 32,
795 & sNy = 32,
796 & OLx = 3,
797 & OLy = 3,
798 & nSx = 6,
799 & nSy = 1,
800 & nPx = 1,
801 & nPy = 1)
802 \end{verbatim}
803 This sets up tiles with x-dimension of thirty-two grid points, y-dimension of
804 thirty-two grid points, and x and y overlaps of three grid points each.
805 There are six tiles allocated to six separate logical processors ({\em nSx=6}).
806 This set of values can be used for a cube sphere calculation.
807 Each tile of size $32 \times 32$ represents a face of the
808 cube. Initializing the tile connectivity correctly ( see section
809 \ref{sect:cube_sphere_communication}. allows the rotations associated with
810 moving between the six cube faces to be embedded within the
811 tile-tile communication code.
812 \end{enumerate}
813
814
815 \subsection{Starting the code}
816 \label{sect:starting_the_code}
817 When code is started under the WRAPPER, execution begins in a main routine {\em
818 eesupp/src/main.F} that is owned by the WRAPPER. Control is transferred
819 to the application through a routine called {\em THE\_MODEL\_MAIN()}
820 once the WRAPPER has initialized correctly and has created the necessary variables
821 to support subsequent calls to communication routines
822 by the application code. The startup calling sequence followed by the
823 WRAPPER is shown in figure \ref{fig:wrapper_startup}.
824
825 \begin{figure}
826 {\footnotesize
827 \begin{verbatim}
828
829 MAIN
830 |
831 |--EEBOOT :: WRAPPER initialization
832 | |
833 | |-- EEBOOT_MINMAL :: Minimal startup. Just enough to
834 | | allow basic I/O.
835 | |-- EEINTRO_MSG :: Write startup greeting.
836 | |
837 | |-- EESET_PARMS :: Set WRAPPER parameters
838 | |
839 | |-- EEWRITE_EEENV :: Print WRAPPER parameter settings
840 | |
841 | |-- INI_PROCS :: Associate processes with grid regions.
842 | |
843 | |-- INI_THREADING_ENVIRONMENT :: Associate threads with grid regions.
844 | |
845 | |--INI_COMMUNICATION_PATTERNS :: Initialize between tile
846 | :: communication data structures
847 |
848 |
849 |--CHECK_THREADS :: Validate multiple thread start up.
850 |
851 |--THE_MODEL_MAIN :: Numerical code top-level driver routine
852
853
854 \end{verbatim}
855 }
856 \caption{Main stages of the WRAPPER startup procedure.
857 This process proceeds transfer of control to application code, which
858 occurs through the procedure {\em THE\_MODEL\_MAIN()}.
859 } \label{fig:wrapper_startup}
860 \end{figure}
861
862 \subsubsection{Multi-threaded execution}
863 \label{sect:multi-threaded-execution}
864 Prior to transferring control to the procedure {\em THE\_MODEL\_MAIN()} the
865 WRAPPER may cause several coarse grain threads to be initialized. The routine
866 {\em THE\_MODEL\_MAIN()} is called once for each thread and is passed a single
867 stack argument which is the thread number, stored in the
868 variable {\em myThid}. In addition to specifying a decomposition with
869 multiple tiles per process ( see section \ref{sect:specifying_a_decomposition})
870 configuring and starting a code to run using multiple threads requires the following
871 steps.\\
872
873 \paragraph{Compilation}
874 First the code must be compiled with appropriate multi-threading directives
875 active in the file {\em main.F} and with appropriate compiler flags
876 to request multi-threading support. The header files
877 {\em MAIN\_PDIRECTIVES1.h} and {\em MAIN\_PDIRECTIVES2.h}
878 contain directives compatible with compilers for Sun, Compaq, SGI,
879 Hewlett-Packard SMP systems and CRAY PVP systems. These directives can be
880 activated by using compile time
881 directives {\em -DTARGET\_SUN},
882 {\em -DTARGET\_DEC}, {\em -DTARGET\_SGI}, {\em -DTARGET\_HP}
883 or {\em -DTARGET\_CRAY\_VECTOR} respectively. Compiler options
884 for invoking multi-threaded compilation vary from system to system
885 and from compiler to compiler. The options will be described
886 in the individual compiler documentation. For the Fortran compiler
887 from Sun the following options are needed to correctly compile
888 multi-threaded code
889 \begin{verbatim}
890 -stackvar -explicitpar -vpara -noautopar
891 \end{verbatim}
892 These options are specific to the Sun compiler. Other compilers
893 will use different syntax that will be described in their
894 documentation. The effect of these options is as follows
895 \begin{enumerate}
896 \item {\bf -stackvar} Causes all local variables to be allocated in stack
897 storage. This is necessary for local variables to ensure that they are private
898 to their thread. Note, when using this option it may be necessary to override
899 the default limit on stack-size that the operating system assigns to a process.
900 This can normally be done by changing the settings of the command shells
901 {\em stack-size} limit variable. However, on some systems changing this limit
902 will require privileged administrator access to modify system parameters.
903
904 \item {\bf -explicitpar} Requests that multiple threads be spawned
905 in response to explicit directives in the application code. These
906 directives are inserted with syntax appropriate to the particular target
907 platform when, for example, the {\em -DTARGET\_SUN} flag is selected.
908
909 \item {\bf -vpara} This causes the compiler to describe the multi-threaded
910 configuration it is creating. This is not required
911 but it can be useful when trouble shooting.
912
913 \item {\bf -noautopar} This inhibits any automatic multi-threaded
914 parallelization the compiler may otherwise generate.
915
916 \end{enumerate}
917
918
919 An example of valid settings for the {\em eedata} file for a
920 domain with two subdomains in y and running with two threads is shown
921 below
922 \begin{verbatim}
923 nTx=1,nTy=2
924 \end{verbatim}
925 This set of values will cause computations to stay within a single
926 thread when moving across the {\em nSx} sub-domains. In the y-direction,
927 however, sub-domains will be split equally between two threads.
928
929 \paragraph{Multi-threading files and parameters} The following
930 files and variables are used in setting up multi-threaded execution.\\
931
932 \fbox{
933 \begin{minipage}{4.75in}
934 File: {\em eesupp/inc/MAIN\_PDIRECTIVES1.h}\\
935 File: {\em eesupp/inc/MAIN\_PDIRECTIVES2.h}\\
936 File: {\em model/src/THE\_MODEL\_MAIN.F}\\
937 File: {\em eesupp/src/MAIN.F}\\
938 File: {\em tools/genmake2}\\
939 File: {\em eedata}\\
940 CPP: {\em TARGET\_SUN}\\
941 CPP: {\em TARGET\_DEC}\\
942 CPP: {\em TARGET\_HP }\\
943 CPP: {\em TARGET\_SGI}\\
944 CPP: {\em TARGET\_CRAY\_VECTOR}\\
945 Parameter: {\em nTx}\\
946 Parameter: {\em nTy}
947 \end{minipage}
948 } \\
949
950 \subsubsection{Multi-process execution}
951 \label{sect:multi-process-execution}
952
953 Despite its appealing programming model, multi-threaded execution remains
954 less common then multi-process execution. One major reason for this
955 is that many system libraries are still not ``thread-safe''. This means that for
956 example on some systems it is not safe to call system routines to
957 do I/O when running in multi-threaded mode, except for in a limited set of
958 circumstances. Another reason is that support for multi-threaded programming
959 models varies between systems.
960
961 Multi-process execution is more ubiquitous.
962 In order to run code in a multi-process configuration a decomposition
963 specification ( see section \ref{sect:specifying_a_decomposition})
964 is given ( in which the at least one of the
965 parameters {\em nPx} or {\em nPy} will be greater than one)
966 and then, as for multi-threaded operation,
967 appropriate compile time and run time steps must be taken.
968
969 \paragraph{Compilation} Multi-process execution under the WRAPPER
970 assumes that the portable, MPI libraries are available
971 for controlling the start-up of multiple processes. The MPI libraries
972 are not required, although they are usually used, for performance
973 critical communication. However, in order to simplify the task
974 of controlling and coordinating the start up of a large number
975 (hundreds and possibly even thousands) of copies of the same
976 program, MPI is used. The calls to the MPI multi-process startup
977 routines must be activated at compile time. Currently MPI libraries are
978 invoked by
979 specifying the appropriate options file with the
980 \begin{verbatim}-of\end{verbatim} flag when running the {\em genmake2}
981 script, which generates the Makefile for compiling and linking MITgcm.
982 (Previously this was done by setting the {\em ALLOW\_USE\_MPI} and
983 {\em ALWAYS\_USE\_MPI} flags in the {\em CPP\_EEOPTIONS.h} file.) More
984 detailed information about the use of {\em genmake2} for specifying
985 local compiler flags is located in section 3 ??\\
986
987
988 \fbox{
989 \begin{minipage}{4.75in}
990 File: {\em tools/genmake2}
991 \end{minipage}
992 } \\
993 \paragraph{\bf Execution} The mechanics of starting a program in
994 multi-process mode under MPI is not standardized. Documentation
995 associated with the distribution of MPI installed on a system will
996 describe how to start a program using that distribution.
997 For the free, open-source MPICH system the MITgcm program is started
998 using a command such as
999 \begin{verbatim}
1000 mpirun -np 64 -machinefile mf ./mitgcmuv
1001 \end{verbatim}
1002 In this example the text {\em -np 64} specifies the number of processes
1003 that will be created. The numeric value {\em 64} must be equal to the
1004 product of the processor grid settings of {\em nPx} and {\em nPy}
1005 in the file {\em SIZE.h}. The parameter {\em mf} specifies that a text file
1006 called ``mf'' will be read to get a list of processor names on
1007 which the sixty-four processes will execute. The syntax of this file
1008 is specified by the MPI distribution
1009 \\
1010
1011 \fbox{
1012 \begin{minipage}{4.75in}
1013 File: {\em SIZE.h}\\
1014 Parameter: {\em nPx}\\
1015 Parameter: {\em nPy}
1016 \end{minipage}
1017 } \\
1018
1019
1020 \paragraph{Environment variables}
1021 On most systems multi-threaded execution also requires the setting
1022 of a special environment variable. On many machines this variable
1023 is called PARALLEL and its values should be set to the number
1024 of parallel threads required. Generally the help pages associated
1025 with the multi-threaded compiler on a machine will explain
1026 how to set the required environment variables for that machines.
1027
1028 \paragraph{Runtime input parameters}
1029 Finally the file {\em eedata} needs to be configured to indicate
1030 the number of threads to be used in the x and y directions.
1031 The variables {\em nTx} and {\em nTy} in this file are used to
1032 specify the information required. The product of {\em nTx} and
1033 {\em nTy} must be equal to the number of threads spawned i.e.
1034 the setting of the environment variable PARALLEL.
1035 The value of {\em nTx} must subdivide the number of sub-domains
1036 in x ({\em nSx}) exactly. The value of {\em nTy} must subdivide the
1037 number of sub-domains in y ({\em nSy}) exactly.
1038 The multiprocess startup of the MITgcm executable {\em mitgcmuv}
1039 is controlled by the routines {\em EEBOOT\_MINIMAL()} and
1040 {\em INI\_PROCS()}. The first routine performs basic steps required
1041 to make sure each process is started and has a textual output
1042 stream associated with it. By default two output files are opened
1043 for each process with names {\bf STDOUT.NNNN} and {\bf STDERR.NNNN}.
1044 The {\bf NNNNN} part of the name is filled in with the process
1045 number so that process number 0 will create output files
1046 {\bf STDOUT.0000} and {\bf STDERR.0000}, process number 1 will create
1047 output files {\bf STDOUT.0001} and {\bf STDERR.0001} etc... These files
1048 are used for reporting status and configuration information and
1049 for reporting error conditions on a process by process basis.
1050 The {\em EEBOOT\_MINIMAL()} procedure also sets the variables
1051 {\em myProcId} and {\em MPI\_COMM\_MODEL}.
1052 These variables are related
1053 to processor identification are are used later in the routine
1054 {\em INI\_PROCS()} to allocate tiles to processes.
1055
1056 Allocation of processes to tiles in controlled by the routine
1057 {\em INI\_PROCS()}. For each process this routine sets
1058 the variables {\em myXGlobalLo} and {\em myYGlobalLo}.
1059 These variables specify (in index space) the coordinate
1060 of the southern most and western most corner of the
1061 southern most and western most tile owned by this process.
1062 The variables {\em pidW}, {\em pidE}, {\em pidS} and {\em pidN}
1063 are also set in this routine. These are used to identify
1064 processes holding tiles to the west, east, south and north
1065 of this process. These values are stored in global storage
1066 in the header file {\em EESUPPORT.h} for use by
1067 communication routines. The above does not hold when the
1068 exch2 package is used -- exch2 sets its own parameters to
1069 specify the global indices of tiles and their relationships
1070 to each other. See exch2 docs for details??.
1071 \\
1072
1073 \fbox{
1074 \begin{minipage}{4.75in}
1075 File: {\em eesupp/src/eeboot\_minimal.F}\\
1076 File: {\em eesupp/src/ini\_procs.F} \\
1077 File: {\em eesupp/inc/EESUPPORT.h} \\
1078 Parameter: {\em myProcId} \\
1079 Parameter: {\em MPI\_COMM\_MODEL} \\
1080 Parameter: {\em myXGlobalLo} \\
1081 Parameter: {\em myYGlobalLo} \\
1082 Parameter: {\em pidW } \\
1083 Parameter: {\em pidE } \\
1084 Parameter: {\em pidS } \\
1085 Parameter: {\em pidN }
1086 \end{minipage}
1087 } \\
1088
1089
1090 \subsection{Controlling communication}
1091 The WRAPPER maintains internal information that is used for communication
1092 operations and that can be customized for different platforms. This section
1093 describes the information that is held and used.
1094
1095 \begin{enumerate}
1096 \item {\bf Tile-tile connectivity information} For each tile the WRAPPER
1097 sets a flag that sets the tile number to the north, south, east and
1098 west of that tile. This number is unique over all tiles in a
1099 configuration. Except when using the exch2 package,
1100 the number is held in the variables {\em tileNo}
1101 ( this holds the tiles own number), {\em tileNoN}, {\em tileNoS},
1102 {\em tileNoE} and {\em tileNoW}. A parameter is also stored with each tile
1103 that specifies the type of communication that is used between tiles.
1104 This information is held in the variables {\em tileCommModeN},
1105 {\em tileCommModeS}, {\em tileCommModeE} and {\em tileCommModeW}.
1106 This latter set of variables can take one of the following values
1107 {\em COMM\_NONE}, {\em COMM\_MSG}, {\em COMM\_PUT} and {\em COMM\_GET}.
1108 A value of {\em COMM\_NONE} is used to indicate that a tile has no
1109 neighbor to communicate with on a particular face. A value
1110 of {\em COMM\_MSG} is used to indicated that some form of distributed
1111 memory communication is required to communicate between
1112 these tile faces ( see section \ref{sect:distributed_memory_communication}).
1113 A value of {\em COMM\_PUT} or {\em COMM\_GET} is used to indicate
1114 forms of shared memory communication ( see section
1115 \ref{sect:shared_memory_communication}). The {\em COMM\_PUT} value indicates
1116 that a CPU should communicate by writing to data structures owned by another
1117 CPU. A {\em COMM\_GET} value indicates that a CPU should communicate by reading
1118 from data structures owned by another CPU. These flags affect the behavior
1119 of the WRAPPER exchange primitive
1120 (see figure \ref{fig:communication_primitives}). The routine
1121 {\em ini\_communication\_patterns()} is responsible for setting the
1122 communication mode values for each tile.
1123 \\
1124
1125 When using the cubed sphere configuration with the exch2 package, the
1126 relationships between tiles and their communication methods are set
1127 by the package in other variables. See the exch2 docs for details.??
1128
1129
1130
1131 \fbox{
1132 \begin{minipage}{4.75in}
1133 File: {\em eesupp/src/ini\_communication\_patterns.F}\\
1134 File: {\em eesupp/inc/EESUPPORT.h} \\
1135 Parameter: {\em tileNo} \\
1136 Parameter: {\em tileNoE} \\
1137 Parameter: {\em tileNoW} \\
1138 Parameter: {\em tileNoN} \\
1139 Parameter: {\em tileNoS} \\
1140 Parameter: {\em tileCommModeE} \\
1141 Parameter: {\em tileCommModeW} \\
1142 Parameter: {\em tileCommModeN} \\
1143 Parameter: {\em tileCommModeS} \\
1144 \end{minipage}
1145 } \\
1146
1147 \item {\bf MP directives}
1148 The WRAPPER transfers control to numerical application code through
1149 the routine {\em THE\_MODEL\_MAIN}. This routine is called in a way
1150 that allows for it to be invoked by several threads. Support for this
1151 is based on using multi-processing (MP) compiler directives.
1152 Most commercially available Fortran compilers support the generation
1153 of code to spawn multiple threads through some form of compiler directives.
1154 As this is generally much more convenient than writing code to interface
1155 to operating system libraries to explicitly spawn threads, and on some systems
1156 this may be the only method available the WRAPPER is distributed with
1157 template MP directives for a number of systems.
1158
1159 These directives are inserted into the code just before and after the
1160 transfer of control to numerical algorithm code through the routine
1161 {\em THE\_MODEL\_MAIN}. Figure \ref{fig:mp_directives} shows an example of
1162 the code that performs this process for a Silicon Graphics system.
1163 This code is extracted from the files {\em main.F} and
1164 {\em MAIN\_PDIRECTIVES1.h}. The variable {\em nThreads} specifies
1165 how many instances of the routine {\em THE\_MODEL\_MAIN} will
1166 be created. The value of {\em nThreads} is set in the routine
1167 {\em INI\_THREADING\_ENVIRONMENT}. The value is set equal to the
1168 the product of the parameters {\em nTx} and {\em nTy} that
1169 are read from the file {\em eedata}. If the value of {\em nThreads}
1170 is inconsistent with the number of threads requested from the
1171 operating system (for example by using an environment
1172 variable as described in section \ref{sect:multi_threaded_execution})
1173 then usually an error will be reported by the routine
1174 {\em CHECK\_THREADS}.\\
1175
1176 \fbox{
1177 \begin{minipage}{4.75in}
1178 File: {\em eesupp/src/ini\_threading\_environment.F}\\
1179 File: {\em eesupp/src/check\_threads.F} \\
1180 File: {\em eesupp/src/main.F} \\
1181 File: {\em eesupp/inc/MAIN\_PDIRECTIVES1.h} \\
1182 File: {\em eedata } \\
1183 Parameter: {\em nThreads} \\
1184 Parameter: {\em nTx} \\
1185 Parameter: {\em nTy} \\
1186 \end{minipage}
1187 }
1188
1189 \item {\bf memsync flags}
1190 As discussed in section \ref{sect:memory_consistency}, when using shared memory,
1191 a low-level system function may be need to force memory consistency.
1192 The routine {\em MEMSYNC()} is used for this purpose. This routine should
1193 not need modifying and the information below is only provided for
1194 completeness. A logical parameter {\em exchNeedsMemSync} set
1195 in the routine {\em INI\_COMMUNICATION\_PATTERNS()} controls whether
1196 the {\em MEMSYNC()} primitive is called. In general this
1197 routine is only used for multi-threaded execution.
1198 The code that goes into the {\em MEMSYNC()}
1199 routine is specific to the compiler and
1200 processor being used for multi-threaded execution and in general
1201 must be written using a short code snippet of assembly language.
1202 For an Ultra Sparc system the following code snippet is used
1203 \begin{verbatim}
1204 asm("membar #LoadStore|#StoreStore");
1205 \end{verbatim}
1206 for an Alpha based system the equivalent code reads
1207 \begin{verbatim}
1208 asm("mb");
1209 \end{verbatim}
1210 while on an x86 system the following code is required
1211 \begin{verbatim}
1212 asm("lock; addl $0,0(%%esp)": : :"memory")
1213 \end{verbatim}
1214
1215 \item {\bf Cache line size}
1216 As discussed in section \ref{sect:cache_effects_and_false_sharing},
1217 milti-threaded codes explicitly avoid penalties associated with excessive
1218 coherence traffic on an SMP system. To do this the shared memory data structures
1219 used by the {\em GLOBAL\_SUM}, {\em GLOBAL\_MAX} and {\em BARRIER} routines
1220 are padded. The variables that control the padding are set in the
1221 header file {\em EEPARAMS.h}. These variables are called
1222 {\em cacheLineSize}, {\em lShare1}, {\em lShare4} and
1223 {\em lShare8}. The default values should not normally need changing.
1224 \item {\bf \_BARRIER}
1225 This is a CPP macro that is expanded to a call to a routine
1226 which synchronizes all the logical processors running under the
1227 WRAPPER. Using a macro here preserves flexibility to insert
1228 a specialized call in-line into application code. By default this
1229 resolves to calling the procedure {\em BARRIER()}. The default
1230 setting for the \_BARRIER macro is given in the file {\em CPP\_EEMACROS.h}.
1231
1232 \item {\bf \_GSUM}
1233 This is a CPP macro that is expanded to a call to a routine
1234 which sums up a floating point number
1235 over all the logical processors running under the
1236 WRAPPER. Using a macro here provides extra flexibility to insert
1237 a specialized call in-line into application code. By default this
1238 resolves to calling the procedure {\em GLOBAL\_SUM\_R8()} ( for
1239 64-bit floating point operands)
1240 or {\em GLOBAL\_SUM\_R4()} (for 32-bit floating point operands). The default
1241 setting for the \_GSUM macro is given in the file {\em CPP\_EEMACROS.h}.
1242 The \_GSUM macro is a performance critical operation, especially for
1243 large processor count, small tile size configurations.
1244 The custom communication example discussed in section \ref{sect:jam_example}
1245 shows how the macro is used to invoke a custom global sum routine
1246 for a specific set of hardware.
1247
1248 \item {\bf \_EXCH}
1249 The \_EXCH CPP macro is used to update tile overlap regions.
1250 It is qualified by a suffix indicating whether overlap updates are for
1251 two-dimensional ( \_EXCH\_XY ) or three dimensional ( \_EXCH\_XYZ )
1252 physical fields and whether fields are 32-bit floating point
1253 ( \_EXCH\_XY\_R4, \_EXCH\_XYZ\_R4 ) or 64-bit floating point
1254 ( \_EXCH\_XY\_R8, \_EXCH\_XYZ\_R8 ). The macro mappings are defined
1255 in the header file {\em CPP\_EEMACROS.h}. As with \_GSUM, the
1256 \_EXCH operation plays a crucial role in scaling to small tile,
1257 large logical and physical processor count configurations.
1258 The example in section \ref{sect:jam_example} discusses defining an
1259 optimized and specialized form on the \_EXCH operation.
1260
1261 The \_EXCH operation is also central to supporting grids such as
1262 the cube-sphere grid. In this class of grid a rotation may be required
1263 between tiles. Aligning the coordinate requiring rotation with the
1264 tile decomposition, allows the coordinate transformation to
1265 be embedded within a custom form of the \_EXCH primitive.
1266
1267 \item {\bf Reverse Mode}
1268 The communication primitives \_EXCH and \_GSUM both employ
1269 hand-written adjoint forms (or reverse mode) forms.
1270 These reverse mode forms can be found in the
1271 source code directory {\em pkg/autodiff}.
1272 For the global sum primitive the reverse mode form
1273 calls are to {\em GLOBAL\_ADSUM\_R4} and
1274 {\em GLOBAL\_ADSUM\_R8}. The reverse mode form of the
1275 exchange primitives are found in routines
1276 prefixed {\em ADEXCH}. The exchange routines make calls to
1277 the same low-level communication primitives as the forward mode
1278 operations. However, the routine argument {\em simulationMode}
1279 is set to the value {\em REVERSE\_SIMULATION}. This signifies
1280 ti the low-level routines that the adjoint forms of the
1281 appropriate communication operation should be performed.
1282 \item {\bf MAX\_NO\_THREADS}
1283 The variable {\em MAX\_NO\_THREADS} is used to indicate the
1284 maximum number of OS threads that a code will use. This
1285 value defaults to thirty-two and is set in the file {\em EEPARAMS.h}.
1286 For single threaded execution it can be reduced to one if required.
1287 The value; is largely private to the WRAPPER and application code
1288 will nor normally reference the value, except in the following scenario.
1289
1290 For certain physical parametrization schemes it is necessary to have
1291 a substantial number of work arrays. Where these arrays are allocated
1292 in heap storage ( for example COMMON blocks ) multi-threaded
1293 execution will require multiple instances of the COMMON block data.
1294 This can be achieved using a Fortran 90 module construct, however,
1295 if this might be unavailable then the work arrays can be extended
1296 with dimensions use the tile dimensioning scheme of {\em nSx}
1297 and {\em nSy} ( as described in section
1298 \ref{sect:specifying_a_decomposition}). However, if the configuration
1299 being specified involves many more tiles than OS threads then
1300 it can save memory resources to reduce the variable
1301 {\em MAX\_NO\_THREADS} to be equal to the actual number of threads that
1302 will be used and to declare the physical parameterization
1303 work arrays with a single {\em MAX\_NO\_THREADS} extra dimension.
1304 An example of this is given in the verification experiment
1305 {\em aim.5l\_cs}. Here the default setting of
1306 {\em MAX\_NO\_THREADS} is altered to
1307 \begin{verbatim}
1308 INTEGER MAX_NO_THREADS
1309 PARAMETER ( MAX_NO_THREADS = 6 )
1310 \end{verbatim}
1311 and several work arrays for storing intermediate calculations are
1312 created with declarations of the form.
1313 \begin{verbatim}
1314 common /FORCIN/ sst1(ngp,MAX_NO_THREADS)
1315 \end{verbatim}
1316 This declaration scheme is not used widely, because most global data
1317 is used for permanent not temporary storage of state information.
1318 In the case of permanent state information this approach cannot be used
1319 because there has to be enough storage allocated for all tiles.
1320 However, the technique can sometimes be a useful scheme for reducing memory
1321 requirements in complex physical parameterizations.
1322 \end{enumerate}
1323
1324 \begin{figure}
1325 \begin{verbatim}
1326 C--
1327 C-- Parallel directives for MIPS Pro Fortran compiler
1328 C--
1329 C Parallel compiler directives for SGI with IRIX
1330 C$PAR PARALLEL DO
1331 C$PAR& CHUNK=1,MP_SCHEDTYPE=INTERLEAVE,
1332 C$PAR& SHARE(nThreads),LOCAL(myThid,I)
1333 C
1334 DO I=1,nThreads
1335 myThid = I
1336
1337 C-- Invoke nThreads instances of the numerical model
1338 CALL THE_MODEL_MAIN(myThid)
1339
1340 ENDDO
1341 \end{verbatim}
1342 \caption{Prior to transferring control to
1343 the procedure {\em THE\_MODEL\_MAIN()} the WRAPPER may use
1344 MP directives to spawn multiple threads.
1345 } \label{fig:mp_directives}
1346 \end{figure}
1347
1348
1349 \subsubsection{Specializing the Communication Code}
1350
1351 The isolation of performance critical communication primitives and the
1352 sub-division of the simulation domain into tiles is a powerful tool.
1353 Here we show how it can be used to improve application performance and
1354 how it can be used to adapt to new griding approaches.
1355
1356 \subsubsection{JAM example}
1357 \label{sect:jam_example}
1358 On some platforms a big performance boost can be obtained by
1359 binding the communication routines {\em \_EXCH} and
1360 {\em \_GSUM} to specialized native libraries ) fro example the
1361 shmem library on CRAY T3E systems). The {\em LETS\_MAKE\_JAM} CPP flag
1362 is used as an illustration of a specialized communication configuration
1363 that substitutes for standard, portable forms of {\em \_EXCH} and
1364 {\em \_GSUM}. It affects three source files {\em eeboot.F},
1365 {\em CPP\_EEMACROS.h} and {\em cg2d.F}. When the flag is defined
1366 is has the following effects.
1367 \begin{itemize}
1368 \item An extra phase is included at boot time to initialize the custom
1369 communications library ( see {\em ini\_jam.F}).
1370 \item The {\em \_GSUM} and {\em \_EXCH} macro definitions are replaced
1371 with calls to custom routines ( see {\em gsum\_jam.F} and {\em exch\_jam.F})
1372 \item a highly specialized form of the exchange operator (optimized
1373 for overlap regions of width one) is substituted into the elliptic
1374 solver routine {\em cg2d.F}.
1375 \end{itemize}
1376 Developing specialized code for other libraries follows a similar
1377 pattern.
1378
1379 \subsubsection{Cube sphere communication}
1380 \label{sect:cube_sphere_communication}
1381 Actual {\em \_EXCH} routine code is generated automatically from
1382 a series of template files, for example {\em exch\_rx.template}.
1383 This is done to allow a large number of variations on the exchange
1384 process to be maintained. One set of variations supports the
1385 cube sphere grid. Support for a cube sphere grid in MITgcm is based
1386 on having each face of the cube as a separate tile (or tiles).
1387 The exchange routines are then able to absorb much of the
1388 detailed rotation and reorientation required when moving around the
1389 cube grid. The set of {\em \_EXCH} routines that contain the
1390 word cube in their name perform these transformations.
1391 They are invoked when the run-time logical parameter
1392 {\em useCubedSphereExchange} is set true. To facilitate the
1393 transformations on a staggered C-grid, exchange operations are defined
1394 separately for both vector and scalar quantities and for
1395 grid-centered and for grid-face and corner quantities.
1396 Three sets of exchange routines are defined. Routines
1397 with names of the form {\em exch\_rx} are used to exchange
1398 cell centered scalar quantities. Routines with names of the form
1399 {\em exch\_uv\_rx} are used to exchange vector quantities located at
1400 the C-grid velocity points. The vector quantities exchanged by the
1401 {\em exch\_uv\_rx} routines can either be signed (for example velocity
1402 components) or un-signed (for example grid-cell separations).
1403 Routines with names of the form {\em exch\_z\_rx} are used to exchange
1404 quantities at the C-grid vorticity point locations.
1405
1406
1407
1408
1409 \section{MITgcm execution under WRAPPER}
1410
1411 Fitting together the WRAPPER elements, package elements and
1412 MITgcm core equation elements of the source code produces calling
1413 sequence shown in section \ref{sect:calling_sequence}
1414
1415 \subsection{Annotated call tree for MITgcm and WRAPPER}
1416 \label{sect:calling_sequence}
1417
1418 WRAPPER layer.
1419
1420 {\footnotesize
1421 \begin{verbatim}
1422
1423 MAIN
1424 |
1425 |--EEBOOT :: WRAPPER initialization
1426 | |
1427 | |-- EEBOOT_MINMAL :: Minimal startup. Just enough to
1428 | | allow basic I/O.
1429 | |-- EEINTRO_MSG :: Write startup greeting.
1430 | |
1431 | |-- EESET_PARMS :: Set WRAPPER parameters
1432 | |
1433 | |-- EEWRITE_EEENV :: Print WRAPPER parameter settings
1434 | |
1435 | |-- INI_PROCS :: Associate processes with grid regions.
1436 | |
1437 | |-- INI_THREADING_ENVIRONMENT :: Associate threads with grid regions.
1438 | |
1439 | |--INI_COMMUNICATION_PATTERNS :: Initialize between tile
1440 | :: communication data structures
1441 |
1442 |
1443 |--CHECK_THREADS :: Validate multiple thread start up.
1444 |
1445 |--THE_MODEL_MAIN :: Numerical code top-level driver routine
1446
1447 \end{verbatim}
1448 }
1449
1450 Core equations plus packages.
1451
1452 {\footnotesize
1453 \begin{verbatim}
1454 C
1455 C
1456 C Invocation from WRAPPER level...
1457 C :
1458 C :
1459 C |
1460 C |-THE_MODEL_MAIN :: Primary driver for the MITgcm algorithm
1461 C | :: Called from WRAPPER level numerical
1462 C | :: code invocation routine. On entry
1463 C | :: to THE_MODEL_MAIN separate thread and
1464 C | :: separate processes will have been established.
1465 C | :: Each thread and process will have a unique ID
1466 C | :: but as yet it will not be associated with a
1467 C | :: specific region in decomposed discrete space.
1468 C |
1469 C |-INITIALISE_FIXED :: Set fixed model arrays such as topography,
1470 C | | :: grid, solver matrices etc..
1471 C | |
1472 C | |-INI_PARMS :: Routine to set kernel model parameters.
1473 C | | :: By default kernel parameters are read from file
1474 C | | :: "data" in directory in which code executes.
1475 C | |
1476 C | |-MON_INIT :: Initializes monitor package ( see pkg/monitor )
1477 C | |
1478 C | |-INI_GRID :: Control grid array (vert. and hori.) initialization.
1479 C | | | :: Grid arrays are held and described in GRID.h.
1480 C | | |
1481 C | | |-INI_VERTICAL_GRID :: Initialize vertical grid arrays.
1482 C | | |
1483 C | | |-INI_CARTESIAN_GRID :: Cartesian horiz. grid initialization
1484 C | | | :: (calculate grid from kernel parameters).
1485 C | | |
1486 C | | |-INI_SPHERICAL_POLAR_GRID :: Spherical polar horiz. grid
1487 C | | | :: initialization (calculate grid from
1488 C | | | :: kernel parameters).
1489 C | | |
1490 C | | |-INI_CURVILINEAR_GRID :: General orthogonal, structured horiz.
1491 C | | :: grid initializations. ( input from raw
1492 C | | :: grid files, LONC.bin, DXF.bin etc... )
1493 C | |
1494 C | |-INI_DEPTHS :: Read (from "bathyFile") or set bathymetry/orgography.
1495 C | |
1496 C | |-INI_MASKS_ETC :: Derive horizontal and vertical cell fractions and
1497 C | | :: land masking for solid-fluid boundaries.
1498 C | |
1499 C | |-INI_LINEAR_PHSURF :: Set ref. surface Bo_surf
1500 C | |
1501 C | |-INI_CORI :: Set coriolis term. zero, f-plane, beta-plane,
1502 C | | :: sphere options are coded.
1503 C | |
1504 C | |-PACAKGES_BOOT :: Start up the optional package environment.
1505 C | | :: Runtime selection of active packages.
1506 C | |
1507 C | |-PACKAGES_READPARMS :: Call active package internal parameter load.
1508 C | | |
1509 C | | |-GMREDI_READPARMS :: GM Package. see pkg/gmredi
1510 C | | |-KPP_READPARMS :: KPP Package. see pkg/kpp
1511 C | | |-SHAP_FILT_READPARMS :: Shapiro filter package. see pkg/shap_filt
1512 C | | |-OBCS_READPARMS :: Open bndy package. see pkg/obcs
1513 C | | |-AIM_READPARMS :: Intermediate Atmos. pacakage. see pkg/aim
1514 C | | |-COST_READPARMS :: Cost function package. see pkg/cost
1515 C | | |-CTRL_INIT :: Control vector support package. see pkg/ctrl
1516 C | | |-OPTIM_READPARMS :: Optimisation support package. see pkg/ctrl
1517 C | | |-GRDCHK_READPARMS :: Gradient check package. see pkg/grdchk
1518 C | | |-ECCO_READPARMS :: ECCO Support Package. see pkg/ecco
1519 C | |
1520 C | |-PACKAGES_CHECK
1521 C | | |
1522 C | | |-KPP_CHECK :: KPP Package. pkg/kpp
1523 C | | |-OBCS_CHECK :: Open bndy Package. pkg/obcs
1524 C | | |-GMREDI_CHECK :: GM Package. pkg/gmredi
1525 C | |
1526 C | |-PACKAGES_INIT_FIXED
1527 C | | |-OBCS_INIT_FIXED :: Open bndy Package. see pkg/obcs
1528 C | | |-FLT_INIT :: Floats Package. see pkg/flt
1529 C | |
1530 C | |-ZONAL_FILT_INIT :: FFT filter Package. see pkg/zonal_filt
1531 C | |
1532 C | |-INI_CG2D :: 2d con. grad solver initialisation.
1533 C | |
1534 C | |-INI_CG3D :: 3d con. grad solver initialisation.
1535 C | |
1536 C | |-CONFIG_SUMMARY :: Provide synopsis of kernel setup.
1537 C | :: Includes annotated table of kernel
1538 C | :: parameter settings.
1539 C |
1540 C |-CTRL_UNPACK :: Control vector support package. see pkg/ctrl
1541 C |
1542 C |-ADTHE_MAIN_LOOP :: Derivative evaluating form of main time stepping loop
1543 C ! :: Auotmatically generated by TAMC/TAF.
1544 C |
1545 C |-CTRL_PACK :: Control vector support package. see pkg/ctrl
1546 C |
1547 C |-GRDCHK_MAIN :: Gradient check package. see pkg/grdchk
1548 C |
1549 C |-THE_MAIN_LOOP :: Main timestepping loop routine.
1550 C | |
1551 C | |-INITIALISE_VARIA :: Set the initial conditions for time evolving
1552 C | | | :: variables
1553 C | | |
1554 C | | |-INI_LINEAR_PHISURF :: Set ref. surface Bo_surf
1555 C | | |
1556 C | | |-INI_CORI :: Set coriolis term. zero, f-plane, beta-plane,
1557 C | | | :: sphere options are coded.
1558 C | | |
1559 C | | |-INI_CG2D :: 2d con. grad solver initialisation.
1560 C | | |-INI_CG3D :: 3d con. grad solver initialisation.
1561 C | | |-INI_MIXING :: Initialise diapycnal diffusivity.
1562 C | | |-INI_DYNVARS :: Initialise to zero all DYNVARS.h arrays (dynamical
1563 C | | | :: fields).
1564 C | | |
1565 C | | |-INI_FIELDS :: Control initializing model fields to non-zero
1566 C | | | |-INI_VEL :: Initialize 3D flow field.
1567 C | | | |-INI_THETA :: Set model initial temperature field.
1568 C | | | |-INI_SALT :: Set model initial salinity field.
1569 C | | | |-INI_PSURF :: Set model initial free-surface height/pressure.
1570 C | | |
1571 C | | |-INI_TR1 :: Set initial tracer 1 distribution.
1572 C | | |
1573 C | | |-THE_CORRECTION_STEP :: Step forward to next time step.
1574 C | | | | :: Here applied to move restart conditions
1575 C | | | | :: (saved in mid timestep) to correct level in
1576 C | | | | :: time (only used for pre-c35).
1577 C | | | |
1578 C | | | |-CALC_GRAD_PHI_SURF :: Return DDx and DDy of surface pressure
1579 C | | | |-CORRECTION_STEP :: Pressure correction to momentum
1580 C | | | |-CYCLE_TRACER :: Move tracers forward in time.
1581 C | | | |-OBCS_APPLY :: Open bndy package. see pkg/obcs
1582 C | | | |-SHAP_FILT_APPLY :: Shapiro filter package. see pkg/shap_filt
1583 C | | | |-ZONAL_FILT_APPLY :: FFT filter package. see pkg/zonal_filt
1584 C | | | |-CONVECTIVE_ADJUSTMENT :: Control static instability mixing.
1585 C | | | | |-FIND_RHO :: Find adjacent densities.
1586 C | | | | |-CONVECT :: Mix static instability.
1587 C | | | | |-TIMEAVE_CUMULATE :: Update convection statistics.
1588 C | | | |
1589 C | | | |-CALC_EXACT_ETA :: Change SSH to flow divergence.
1590 C | | |
1591 C | | |-CONVECTIVE_ADJUSTMENT_INI :: Control static instability mixing
1592 C | | | | :: Extra time history interactions.
1593 C | | | |
1594 C | | | |-FIND_RHO :: Find adjacent densities.
1595 C | | | |-CONVECT :: Mix static instability.
1596 C | | | |-TIMEAVE_CUMULATE :: Update convection statistics.
1597 C | | |
1598 C | | |-PACKAGES_INIT_VARIABLES :: Does initialisation of time evolving
1599 C | | | | :: package data.
1600 C | | | |
1601 C | | | |-GMREDI_INIT :: GM package. ( see pkg/gmredi )
1602 C | | | |-KPP_INIT :: KPP package. ( see pkg/kpp )
1603 C | | | |-KPP_OPEN_DIAGS
1604 C | | | |-OBCS_INIT_VARIABLES :: Open bndy. package. ( see pkg/obcs )
1605 C | | | |-AIM_INIT :: Interm. atmos package. ( see pkg/aim )
1606 C | | | |-CTRL_MAP_INI :: Control vector package.( see pkg/ctrl )
1607 C | | | |-COST_INIT :: Cost function package. ( see pkg/cost )
1608 C | | | |-ECCO_INIT :: ECCO support package. ( see pkg/ecco )
1609 C | | | |-INI_FORCING :: Set model initial forcing fields.
1610 C | | | | :: Either set in-line or from file as shown.
1611 C | | | |-READ_FLD_XY_RS(zonalWindFile)
1612 C | | | |-READ_FLD_XY_RS(meridWindFile)
1613 C | | | |-READ_FLD_XY_RS(surfQFile)
1614 C | | | |-READ_FLD_XY_RS(EmPmRfile)
1615 C | | | |-READ_FLD_XY_RS(thetaClimFile)
1616 C | | | |-READ_FLD_XY_RS(saltClimFile)
1617 C | | | |-READ_FLD_XY_RS(surfQswFile)
1618 C | | |
1619 C | | |-CALC_SURF_DR :: Calculate the new surface level thickness.
1620 C | | |-UPDATE_SURF_DR :: Update the surface-level thickness fraction.
1621 C | | |-UPDATE_CG2D :: Update 2d conjugate grad. for Free-Surf.
1622 C | | |-STATE_SUMMARY :: Summarize model prognostic variables.
1623 C | | |-TIMEAVE_STATVARS :: Time averaging package ( see pkg/timeave ).
1624 C | |
1625 C | |-WRITE_STATE :: Controlling routine for IO to dump model state.
1626 C | | |-WRITE_REC_XYZ_RL :: Single file I/O
1627 C | | |-WRITE_FLD_XYZ_RL :: Multi-file I/O
1628 C | |
1629 C | |-MONITOR :: Monitor state ( see pkg/monitor )
1630 C | |-CTRL_MAP_FORCING :: Control vector support package. ( see pkg/ctrl )
1631 C====|>|
1632 C====|>| ****************************
1633 C====|>| BEGIN MAIN TIMESTEPPING LOOP
1634 C====|>| ****************************
1635 C====|>|
1636 C/\ | |-FORWARD_STEP :: Step forward a time-step ( AT LAST !!! )
1637 C/\ | | |
1638 C/\ | | |-DUMMY_IN_STEPPING :: autodiff package ( pkg/autoduff ).
1639 C/\ | | |-CALC_EXACT_ETA :: Change SSH to flow divergence.
1640 C/\ | | |-CALC_SURF_DR :: Calculate the new surface level thickness.
1641 C/\ | | |-EXF_GETFORCING :: External forcing package. ( pkg/exf )
1642 C/\ | | |-EXTERNAL_FIELDS_LOAD :: Control loading time dep. external data.
1643 C/\ | | | | :: Simple interpolation between end-points
1644 C/\ | | | | :: for forcing datasets.
1645 C/\ | | | |
1646 C/\ | | | |-EXCH :: Sync forcing. in overlap regions.
1647 C/\ | | |
1648 C/\ | | |-THERMODYNAMICS :: theta, salt + tracer equations driver.
1649 C/\ | | | |
1650 C/\ | | | |-INTEGRATE_FOR_W :: Integrate for vertical velocity.
1651 C/\ | | | |-OBCS_APPLY_W :: Open bndy. package ( see pkg/obcs ).
1652 C/\ | | | |-FIND_RHO :: Calculates [rho(S,T,z)-Rhonil] of a slice
1653 C/\ | | | |-GRAD_SIGMA :: Calculate isoneutral gradients
1654 C/\ | | | |-CALC_IVDC :: Set Implicit Vertical Diffusivity for Convection
1655 C/\ | | | |
1656 C/\ | | | |-OBCS_CALC :: Open bndy. package ( see pkg/obcs ).
1657 C/\ | | | |-EXTERNAL_FORCING_SURF:: Accumulates appropriately dimensioned
1658 C/\ | | | | :: forcing terms.
1659 C/\ | | | |
1660 C/\ | | | |-GMREDI_CALC_TENSOR :: GM package ( see pkg/gmredi ).
1661 C/\ | | | |-GMREDI_CALC_TENSOR_DUMMY :: GM package ( see pkg/gmredi ).
1662 C/\ | | | |-KPP_CALC :: KPP package ( see pkg/kpp ).
1663 C/\ | | | |-KPP_CALC_DUMMY :: KPP package ( see pkg/kpp ).
1664 C/\ | | | |-AIM_DO_ATMOS_PHYSICS :: Intermed. atmos package ( see pkg/aim ).
1665 C/\ | | | |-GAD_ADVECTION :: Generalised advection driver (multi-dim
1666 C/\ | | | | advection case) (see pkg/gad).
1667 C/\ | | | |-CALC_COMMON_FACTORS :: Calculate common data (such as volume flux)
1668 C/\ | | | |-CALC_DIFFUSIVITY :: Calculate net vertical diffusivity
1669 C/\ | | | | |
1670 C/\ | | | | |-GMREDI_CALC_DIFF :: GM package ( see pkg/gmredi ).
1671 C/\ | | | | |-KPP_CALC_DIFF :: KPP package ( see pkg/kpp ).
1672 C/\ | | | |
1673 C/\ | | | |-CALC_GT :: Calculate the temperature tendency terms
1674 C/\ | | | | |
1675 C/\ | | | | |-GAD_CALC_RHS :: Generalised advection package
1676 C/\ | | | | | :: ( see pkg/gad )
1677 C/\ | | | | |-EXTERNAL_FORCING_T :: Problem specific forcing for temperature.
1678 C/\ | | | | |-ADAMS_BASHFORTH2 :: Extrapolate tendencies forward in time.
1679 C/\ | | | | |-FREESURF_RESCALE_G :: Re-scale Gt for free-surface height.
1680 C/\ | | | |
1681 C/\ | | | |-TIMESTEP_TRACER :: Step tracer field forward in time
1682 C/\ | | | |
1683 C/\ | | | |-CALC_GS :: Calculate the salinity tendency terms
1684 C/\ | | | | |
1685 C/\ | | | | |-GAD_CALC_RHS :: Generalised advection package
1686 C/\ | | | | | :: ( see pkg/gad )
1687 C/\ | | | | |-EXTERNAL_FORCING_S :: Problem specific forcing for salt.
1688 C/\ | | | | |-ADAMS_BASHFORTH2 :: Extrapolate tendencies forward in time.
1689 C/\ | | | | |-FREESURF_RESCALE_G :: Re-scale Gs for free-surface height.
1690 C/\ | | | |
1691 C/\ | | | |-TIMESTEP_TRACER :: Step tracer field forward in time
1692 C/\ | | | |
1693 C/\ | | | |-CALC_GTR1 :: Calculate other tracer(s) tendency terms
1694 C/\ | | | | |
1695 C/\ | | | | |-GAD_CALC_RHS :: Generalised advection package
1696 C/\ | | | | | :: ( see pkg/gad )
1697 C/\ | | | | |-EXTERNAL_FORCING_TR:: Problem specific forcing for tracer.
1698 C/\ | | | | |-ADAMS_BASHFORTH2 :: Extrapolate tendencies forward in time.
1699 C/\ | | | | |-FREESURF_RESCALE_G :: Re-scale Gs for free-surface height.
1700 C/\ | | | |
1701 C/\ | | | |-TIMESTEP_TRACER :: Step tracer field forward in time
1702 C/\ | | | |-OBCS_APPLY_TS :: Open bndy. package (see pkg/obcs ).
1703 C/\ | | | |-FREEZE :: Limit range of temperature.
1704 C/\ | | | |
1705 C/\ | | | |-IMPLDIFF :: Solve vertical implicit diffusion equation.
1706 C/\ | | | |-OBCS_APPLY_TS :: Open bndy. package (see pkg/obcs ).
1707 C/\ | | | |
1708 C/\ | | | |-AIM_AIM2DYN_EXCHANGES :: Inetermed. atmos (see pkg/aim).
1709 C/\ | | | |-EXCH :: Update overlaps
1710 C/\ | | |
1711 C/\ | | |-DYNAMICS :: Momentum equations driver.
1712 C/\ | | | |
1713 C/\ | | | |-CALC_GRAD_PHI_SURF :: Calculate the gradient of the surface
1714 C/\ | | | | Potential anomaly.
1715 C/\ | | | |-CALC_VISCOSITY :: Calculate net vertical viscosity
1716 C/\ | | | | |-KPP_CALC_VISC :: KPP package ( see pkg/kpp ).
1717 C/\ | | | |
1718 C/\ | | | |-CALC_PHI_HYD :: Integrate the hydrostatic relation.
1719 C/\ | | | |-MOM_FLUXFORM :: Flux form mom eqn. package ( see
1720 C/\ | | | | pkg/mom_fluxform ).
1721 C/\ | | | |-MOM_VECINV :: Vector invariant form mom eqn. package ( see
1722 C/\ | | | | pkg/mom_vecinv ).
1723 C/\ | | | |-TIMESTEP :: Step momentum fields forward in time
1724 C/\ | | | |-OBCS_APPLY_UV :: Open bndy. package (see pkg/obcs ).
1725 C/\ | | | |
1726 C/\ | | | |-IMPLDIFF :: Solve vertical implicit diffusion equation.
1727 C/\ | | | |-OBCS_APPLY_UV :: Open bndy. package (see pkg/obcs ).
1728 C/\ | | | |
1729 C/\ | | | |-TIMEAVE_CUMUL_1T :: Time averaging package ( see pkg/timeave ).
1730 C/\ | | | |-TIMEAVE_CUMUATE :: Time averaging package ( see pkg/timeave ).
1731 C/\ | | | |-DEBUG_STATS_RL :: Quick debug package ( see pkg/debug ).
1732 C/\ | | |
1733 C/\ | | |-CALC_GW :: vert. momentum tendency terms ( NH, QH only ).
1734 C/\ | | |
1735 C/\ | | |-UPDATE_SURF_DR :: Update the surface-level thickness fraction.
1736 C/\ | | |
1737 C/\ | | |-UPDATE_CG2D :: Update 2d conjugate grad. for Free-Surf.
1738 C/\ | | |
1739 C/\ | | |-SOLVE_FOR_PRESSURE :: Find surface pressure.
1740 C/\ | | | |-CALC_DIV_GHAT :: Form the RHS of the surface pressure eqn.
1741 C/\ | | | |-CG2D :: Two-dim pre-con. conjugate-gradient.
1742 C/\ | | | |-CG3D :: Three-dim pre-con. conjugate-gradient solver.
1743 C/\ | | |
1744 C/\ | | |-THE_CORRECTION_STEP :: Step forward to next time step.
1745 C/\ | | | |
1746 C/\ | | | |-CALC_GRAD_PHI_SURF :: Return DDx and DDy of surface pressure
1747 C/\ | | | |-CORRECTION_STEP :: Pressure correction to momentum
1748 C/\ | | | |-CYCLE_TRACER :: Move tracers forward in time.
1749 C/\ | | | |-OBCS_APPLY :: Open bndy package. see pkg/obcs
1750 C/\ | | | |-SHAP_FILT_APPLY :: Shapiro filter package. see pkg/shap_filt
1751 C/\ | | | |-ZONAL_FILT_APPLY :: FFT filter package. see pkg/zonal_filt
1752 C/\ | | | |-CONVECTIVE_ADJUSTMENT :: Control static instability mixing.
1753 C/\ | | | | |-FIND_RHO :: Find adjacent densities.
1754 C/\ | | | | |-CONVECT :: Mix static instability.
1755 C/\ | | | | |-TIMEAVE_CUMULATE :: Update convection statistics.
1756 C/\ | | | |
1757 C/\ | | | |-CALC_EXACT_ETA :: Change SSH to flow divergence.
1758 C/\ | | |
1759 C/\ | | |-DO_FIELDS_BLOCKING_EXCHANGES :: Sync up overlap regions.
1760 C/\ | | | |-EXCH
1761 C/\ | | |
1762 C/\ | | |-FLT_MAIN :: Float package ( pkg/flt ).
1763 C/\ | | |
1764 C/\ | | |-MONITOR :: Monitor package ( pkg/monitor ).
1765 C/\ | | |
1766 C/\ | | |-DO_THE_MODEL_IO :: Standard diagnostic I/O.
1767 C/\ | | | |-WRITE_STATE :: Core state I/O
1768 C/\ | | | |-TIMEAVE_STATV_WRITE :: Time averages. see pkg/timeave
1769 C/\ | | | |-AIM_WRITE_DIAGS :: Intermed. atmos diags. see pkg/aim
1770 C/\ | | | |-GMREDI_DIAGS :: GM diags. see pkg/gmredi
1771 C/\ | | | |-KPP_DO_DIAGS :: KPP diags. see pkg/kpp
1772 C/\ | | |
1773 C/\ | | |-WRITE_CHECKPOINT :: Do I/O for restart files.
1774 C/\ | |
1775 C/\ | |-COST_TILE :: Cost function package. ( see pkg/cost )
1776 C<===|=|
1777 C<===|=| **************************
1778 C<===|=| END MAIN TIMESTEPPING LOOP
1779 C<===|=| **************************
1780 C<===|=|
1781 C | |-COST_FINAL :: Cost function package. ( see pkg/cost )
1782 C |
1783 C |-WRITE_CHECKPOINT :: Final state storage, for restart.
1784 C |
1785 C |-TIMER_PRINTALL :: Computational timing summary
1786 C |
1787 C |-COMM_STATS :: Summarise inter-proc and inter-thread communication
1788 C :: events.
1789 C
1790 \end{verbatim}
1791 }
1792
1793 \subsection{Measuring and Characterizing Performance}
1794
1795 TO BE DONE (CNH)
1796
1797 \subsection{Estimating Resource Requirements}
1798
1799 TO BE DONE (CNH)
1800
1801 \subsubsection{Atlantic 1/6 degree example}
1802 \subsubsection{Dry Run testing}
1803 \subsubsection{Adjoint Resource Requirements}
1804 \subsubsection{State Estimation Environment Resources}
1805

  ViewVC Help
Powered by ViewVC 1.1.22