multi

1 ############################################################################### 2 # # 3 # Copyright (C) 2007 Gary S Thompson (https://gna.org/users/varioustoxins) # 4 # Copyright (C) 2011-2012 Edward d'Auvergne # 5 # # 6 # This file is part of the program relax. # 7 # # 8 # relax is free software; you can redistribute it and/or modify # 9 # it under the terms of the GNU General Public License as published by # 10 # the Free Software Foundation; either version 2 of the License, or # 11 # (at your option) any later version. # 12 # # 13 # relax is distributed in the hope that it will be useful, # 14 # but WITHOUT ANY WARRANTY; without even the implied warranty of # 15 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # 16 # GNU General Public License for more details. # 17 # # 18 # You should have received a copy of the GNU General Public License # 19 # along with relax; if not, write to the Free Software # 20 # Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA # 21 # # 22 ############################################################################### 23 24 # Package docstring. 25 """The multi-processor package. 26 27 1 Introduction 28 ============== 29 30 This package is an abstraction of specific multi-processor implementations or fabrics such as MPI via mpi4py. It is designed to be extended for use on other fabrics such as grid computing via SSH tunnelling, threading, etc. It also has a uni-processor mode as the default fabric. 31 32 33 2 API 34 ===== 35 36 The public API is available via the __init__ module. It consists of a number of functions and classes. Using this basic interface, code can be parallelised and executed via an MPI implementation, or default back to a single CPU when needed. The choice of processor fabric is up to the calling program (via multi.load_multiprocessor). 37 38 39 2.1 Program initialisation 40 -------------------------- 41 42 The function multi.load_multiprocessor() is the interface for how a program can load and set up a specific processor fabric. This function returns the set up processor, which itself provides a run() method which is used to execute your application. 43 44 45 2.2 Access to the processor instance 46 ------------------------------------ 47 48 The multi.Processor_box class is a special singleton object which provides access to the processor object. This is required for a number of actions: 49 50 - Queuing of slave commands and memos via Processor_box().processor.add_to_queue(). 51 - Returning results (as a Results_command) from the slave processor to the master via Processor_box().processor.return_object(). 52 - Determining the number of processes via Processor_box().processor.processor_size(). 53 - Waiting for completion of the queued slave processors via Processor_box().processor.run_queue(). 54 55 56 2.3 Slaves 57 ---------- 58 59 Slave processors are created via the multi.Slave_command class. This is special base class which must be subclassed. The run() function should be overridden, this provides the code to execute on the slave processors. 60 61 62 2.4 Results handling 63 -------------------- 64 65 The multi.Result_command class is a special base class which must be subclassed. The run() function should be overridden, this provides the code for the master to process the results from the slaves. 66 67 In addition, the multi.Memo should also be used. This is a special base class which must be subclassed. This is a data store used by the Results_command to help process the results from the slave on the master processor. 68 69 70 71 3 Parallelisation 72 ================= 73 74 The following are the steps required to parallelise a calculation via the multi-processor package API. It is assumed that the multi.load_multiprocessor() function has been set up at the highest level so that the entire program will be executed by the returned processor's run() method. 75 76 77 3.1 Subclassing command and memo objects 78 ---------------------------------------- 79 80 The first step is that the Slave_command, Result_command, and Memo classes need to be subclassed. The Slave_command.run() method must be provided and is used for running the calculations on the slave processors. The Result_command is used to unpack the results from the slave. It is initialised by the Slave_command itself with the results from the calculation as arguments of __init__(). Its run() method processes the results on the master processor. The Memo object holds data other than the calculation results required by the Result_command.run() method to process the results. 81 82 83 3.2 Initialisation and queuing 84 ------------------------------ 85 86 The second step is to initialise the Slave_command and Memo and add these to the processor queue. But first access to the processor is required. The singleton multi.Processor_box should be imported, and the processor accessed with code such as:: 87 88 # Initialise the Processor box singleton. 89 processor_box = Processor_box() 90 91 The slave command is then initialised and all required data by the slave for the calculation (via its run() method) is stored within the class instance. The memo is also initialised with its data required for the result command for processing on the master of the results from the slave. These are then queued on the processor:: 92 93 # Queue the slave command and memo. 94 processor_box.processor.add_to_queue(command, memo) 95 96 97 3.3 Calculation 98 --------------- 99 100 To execute the calculations, the final part of the calculation code on the master must feature a call to:: 101 102 processor_box.processor.run_queue(). 103 104 105 4 Example 106 ========= 107 108 See the script 'test_implementation.py' for a basic example of a reference, and full, implementation of the multi-processor package. 109 110 111 5 Issues 112 ======== 113 114 For multi-core systems and Linux 2.6, the following might be required to prevent the master processor from taking 100% of one CPU core while waiting for the slaves: 115 116 # echo "1" > /proc/sys/kernel/sched_compat_yield 117 118 This appears to be an OpenMPI problem with late 2.6 Linux kernels. 119 """ 120 121 122 __all__ = ['memo', 123 'misc', 124 'mpi4py_processor', 125 'multi_processor_base', 126 'processor', 127 'processor_io', 128 'result_commands', 129 'result_queue', 130 'slave_commands', 131 'uni_processor'] 132 133 # Python module imports. 134 import sys as _sys 135 import traceback as _traceback 136 137 # Multi-processor module imports. 138 from multi.memo import Memo 139 from multi.misc import import_module as _import_module 140 from multi.misc import Verbosity as _Verbosity; _verbosity = _Verbosity() 141 from multi.result_commands import Result_command 142 from multi.slave_commands import Slave_command 143 144 145 #FIXME error checking for if module required not found. 146 #FIXME module loading code needs to be in a util module. 147 #FIXME: remove parameters that are not required to load the module (processor_size).

148 -def load_multiprocessor(processor_name, callback, processor_size, verbosity=1):

149 """Load a multi processor given its name. 150 151 Dynamically load a multi processor, the current algorithm is to search in module multi for a 152 module called <processor_name>.<Processor_name> (note capitalisation). 153 154 155 @todo: This algorithm needs to be improved to allow users to load processors without altering the relax source code. 156 157 @todo: Remove non-essential parameters. 158 159 @param processor_name: Name of the processor module/class to load. 160 @type processor_name: str 161 @keyword verbosity: The verbosity level at initialisation. This can be changed during program execution. A value of 0 suppresses all output. A value of 1 causes the basic multi-processor information to be printed. A value of 2 will switch on a number of debugging printouts. Values greater than 2 currently do nothing, though this might change in the future. 162 @type verbosity: int 163 @return: A loaded processor object or None to indicate failure. 164 @rtype: multi.processor.Processor instance 165 """ 166 167 # Check that the processor type is supported. 168 if processor_name not in ['uni', 'mpi4py']: 169 _sys.stderr.write("The processor type '%s' is not supported.\n" % processor_name) 170 _sys.exit() 171 172 # Store the verbosity level. 173 _verbosity.set(verbosity) 174 175 # The Processor details. 176 processor_name = processor_name + '_processor' 177 class_name = processor_name[0].upper() + processor_name[1:] 178 module_path = '.'.join(('multi', processor_name)) 179 180 # Load the module containing the specific processor. 181 modules = _import_module(module_path) 182 183 # Access the class from within the module. 184 if hasattr(modules[-1], class_name): 185 clazz = getattr(modules[-1], class_name) 186 else: 187 raise Exception("can't load class %s from module %s" % (class_name, module_path)) 188 189 # Instantiate the Processor. 190 object = clazz(callback=callback, processor_size=processor_size) 191 192 # Load the Processor_box container and store the details and Processor instance. 193 processor_box = Processor_box() 194 processor_box.processor = object 195 processor_box.processor_name = processor_name 196 processor_box.class_name = class_name 197 198 # Return the Processor instance. 199 return object

200 201

202 -def fetch_data(name=None):

203 """API function for obtaining data from the Processor instance's data store. 204 205 This is for fetching data from the data store of the Processor instance. If run on the master, then the master's data store will be accessed. If run on the slave, then the slave's data store will be accessed. 206 207 208 @attention: No inter-processor communications are performed. 209 210 @keyword name: The name of the data structure to fetch. 211 @type name: str 212 @return: The value of the associated data structure. 213 @rtype: anything 214 """ 215 216 # Load the Processor_box. 217 processor_box = Processor_box() 218 219 # Forward the call to the processor instance. 220 return processor_box.processor.fetch_data(name=name)

221 222

223 -def fetch_data_store():

224 """API function for obtaining the data store object from the Processor instance. 225 226 If run on the master, then the master's data store will be returned. If run on the slave, then the slave's data store will be returned. 227 228 229 @attention: No inter-processor communications are performed. 230 231 @return: The data store of the processor (of the same rank as the calling code). 232 @rtype: class instance 233 """ 234 235 # Load the Processor_box. 236 processor_box = Processor_box() 237 238 # Return the data store. 239 return processor_box.processor.data_store

240 241

242 -def send_data_to_slaves(name=None, value=None):

243 """API function for sending data from the master to all slaves processors. 244 245 @attention: Inter-processor communications are performed. 246 247 @keyword name: The name of the data structure to store. 248 @type name: str 249 @keyword value: The data structure. 250 @type value: anything 251 """ 252 253 # Load the Processor_box. 254 processor_box = Processor_box() 255 256 # Forward the call to the processor instance. 257 processor_box.processor.send_data_to_slaves(name=name, value=value)

258 259 260

261 -class Application_callback(object):

262 """Call backs provided to the host application by the multi processor framework. 263 264 This class allows for independence from the host class/application. 265 266 @note: B{The logic behind the design} the callbacks are defined as two attributes 267 self.init_master and self.handle_exception as handle_exception can be null (which is 268 used to request the use of the processors default error handling code). Note, however, 269 that a class with the equivalent methods would also works as python effectively handles 270 methods as attributes of a class. The signatures for the callback methods are documented 271 by the default methods default_init_master & default_handle_exception. 272 """ 273

274 - def __init__(self, master):

275 """Initialise the callback interface. 276 277 @param master: The data for the host application. In the default implementation this is an 278 object we call methods on but it could be anything... 279 @type master: object 280 """ 281 282 self.master = master 283 """The host application.""" 284 285 self.init_master = self.default_init_master 286 self.handle_exception = self.default_handle_exception

287 288

289 - def default_handle_exception(self, processor, exception):

290 """Handle an exception raised in the processor framework. 291 292 The function is responsible for aborting the processor by calling processor.abort() as its 293 final act. 294 295 @param processor: The processor instance. 296 @type processor: multi.processor.Processor instance 297 @param exception: The exception raised by the processor or slave processor. In the case of 298 a slave processor exception this may well be a wrapped exception of type 299 multi.processor.Capturing_exception which was raised at the point the 300 exception was received on the master processor but contains an enclosed 301 exception from a slave. 302 @type exception: Exception instance 303 """ 304 305 # Print the traceback. 306 _traceback.print_exc(file=_sys.stderr) 307 308 # Stop the processor. 309 processor.abort()

310 311

312 - def default_init_master(self, processor):

313 """Start the main loop of the host application. 314 315 @param processor: The processor instance. 316 @type processor: multi.processor.Processor instance 317 """ 318 319 self.master.run()

320 321 322

323 -class Processor_box(object):

324 """A storage class for the Processor instance and its attributes. 325 326 This singleton contains Processor instances and information about these Processors. Importantly 327 this container gives the calling code access to the Processor. 328 """ 329 330 # Class variable for storing the class instance. 331 instance = None 332

333 - def __new__(self, *args, **kargs):

334 """Replacement function for implementing the singleton design pattern.""" 335 336 # First initialisation. 337 if self.instance is None: 338 self.instance = object.__new__(self, *args, **kargs) 339 340 # Already initialised, so return the instance. 341 return self.instance

342

Source Code for Package multi