multi

1 ############################################################################### 2 # # 3 # Copyright (C) 2007 Gary S Thompson # 4 # Copyright (C) 2008,2012 Edward d'Auvergne # 5 # # 6 # This file is part of the program relax (http://www.nmr-relax.com). # 7 # # 8 # This program 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 3 of the License, or # 11 # (at your option) any later version. # 12 # # 13 # This program 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 this program. If not, see <http://www.gnu.org/licenses/>. # 20 # # 21 ############################################################################### 22 23 # Package docstring. 24 """The multi-processor package. 25 26 1 Introduction 27 ============== 28 29 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. 30 31 32 2 API 33 ===== 34 35 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). 36 37 38 2.1 Program initialisation 39 -------------------------- 40 41 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. 42 43 44 2.2 Access to the processor instance 45 ------------------------------------ 46 47 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: 48 49 - Queuing of slave commands and memos via Processor_box().processor.add_to_queue(). 50 - Returning results (as a Results_command) from the slave processor to the master via Processor_box().processor.return_object(). 51 - Determining the number of processes via Processor_box().processor.processor_size(). 52 - Waiting for completion of the queued slave processors via Processor_box().processor.run_queue(). 53 54 55 2.3 Slaves 56 ---------- 57 58 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. 59 60 61 2.4 Results handling 62 -------------------- 63 64 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. 65 66 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. 67 68 69 70 3 Parallelisation 71 ================= 72 73 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. 74 75 76 3.1 Subclassing command and memo objects 77 ---------------------------------------- 78 79 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. 80 81 82 3.2 Initialisation and queuing 83 ------------------------------ 84 85 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:: 86 87 # Initialise the Processor box singleton. 88 processor_box = Processor_box() 89 90 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:: 91 92 # Queue the slave command and memo. 93 processor_box.processor.add_to_queue(command, memo) 94 95 96 3.3 Calculation 97 --------------- 98 99 To execute the calculations, the final part of the calculation code on the master must feature a call to:: 100 101 processor_box.processor.run_queue(). 102 103 104 4 Example 105 ========= 106 107 See the script 'test_implementation.py' for a basic example of a reference, and full, implementation of the multi-processor package. 108 109 110 5 Issues 111 ======== 112 113 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: 114 115 # echo "1" > /proc/sys/kernel/sched_compat_yield 116 117 This appears to be an OpenMPI problem with late 2.6 Linux kernels. 118 """ 119 120 121 __all__ = ['memo', 122 'misc', 123 'mpi4py_processor', 124 'multi_processor_base', 125 'processor', 126 'processor_io', 127 'result_commands', 128 'result_queue', 129 'slave_commands', 130 'uni_processor'] 131 132 # Python module imports. 133 import sys as _sys 134 import traceback as _traceback 135 136 # Multi-processor module imports. 137 from multi.memo import Memo 138 from multi.misc import import_module as _import_module 139 from multi.misc import Verbosity as _Verbosity; _verbosity = _Verbosity() 140 from multi.result_commands import Result_command 141 from multi.slave_commands import Slave_command 142 143 144 #FIXME error checking for if module required not found. 145 #FIXME module loading code needs to be in a util module. 146 #FIXME: remove parameters that are not required to load the module (processor_size).

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

148 """Load a multi processor given its name. 149 150 Dynamically load a multi processor, the current algorithm is to search in module multi for a 151 module called <processor_name>.<Processor_name> (note capitalisation). 152 153 154 @todo: This algorithm needs to be improved to allow users to load processors without altering the relax source code. 155 156 @todo: Remove non-essential parameters. 157 158 @param processor_name: Name of the processor module/class to load. 159 @type processor_name: str 160 @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. 161 @type verbosity: int 162 @return: A loaded processor object or None to indicate failure. 163 @rtype: multi.processor.Processor instance 164 """ 165 166 # Store the verbosity level. 167 _verbosity.set(verbosity) 168 169 # The Processor details. 170 processor_name = processor_name + '_processor' 171 class_name = processor_name[0].upper() + processor_name[1:] 172 module_path = '.'.join(('multi', processor_name)) 173 174 # Load the module containing the specific processor. 175 modules = _import_module(module_path) 176 177 # Access the class from within the module. 178 if hasattr(modules[-1], class_name): 179 clazz = getattr(modules[-1], class_name) 180 else: 181 raise Exception("can't load class %s from module %s" % (class_name, module_path)) 182 183 # Instantiate the Processor. 184 object = clazz(callback=callback, processor_size=processor_size) 185 186 # Load the Processor_box container and store the details and Processor instance. 187 processor_box = Processor_box() 188 processor_box.processor = object 189 processor_box.processor_name = processor_name 190 processor_box.class_name = class_name 191 192 # Return the Processor instance. 193 return object

194 195

196 -def fetch_data(name=None):

197 """API function for obtaining data from the Processor instance's data store. 198 199 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. 200 201 202 @attention: No inter-processor communications are performed. 203 204 @keyword name: The name of the data structure to fetch. 205 @type name: str 206 @return: The value of the associated data structure. 207 @rtype: anything 208 """ 209 210 # Load the Processor_box. 211 processor_box = Processor_box() 212 213 # Forward the call to the processor instance. 214 return processor_box.processor.fetch_data(name=name)

215 216

217 -def fetch_data_store():

218 """API function for obtaining the data store object from the Processor instance. 219 220 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. 221 222 223 @attention: No inter-processor communications are performed. 224 225 @return: The data store of the processor (of the same rank as the calling code). 226 @rtype: class instance 227 """ 228 229 # Load the Processor_box. 230 processor_box = Processor_box() 231 232 # Return the data store. 233 return processor_box.processor.data_store

234 235

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

237 """API function for sending data from the master to all slaves processors. 238 239 @attention: Inter-processor communications are performed. 240 241 @keyword name: The name of the data structure to store. 242 @type name: str 243 @keyword value: The data structure. 244 @type value: anything 245 """ 246 247 # Load the Processor_box. 248 processor_box = Processor_box() 249 250 # Forward the call to the processor instance. 251 processor_box.processor.send_data_to_slaves(name=name, value=value)

252 253 254

255 -class Application_callback(object):

256 """Call backs provided to the host application by the multi processor framework. 257 258 This class allows for independence from the host class/application. 259 260 @note: B{The logic behind the design} the callbacks are defined as two attributes 261 self.init_master and self.handle_exception as handle_exception can be null (which is 262 used to request the use of the processors default error handling code). Note, however, 263 that a class with the equivalent methods would also works as python effectively handles 264 methods as attributes of a class. The signatures for the callback methods are documented 265 by the default methods default_init_master & default_handle_exception. 266 """ 267

268 - def __init__(self, master):

269 """Initialise the callback interface. 270 271 @param master: The data for the host application. In the default implementation this is an 272 object we call methods on but it could be anything... 273 @type master: object 274 """ 275 276 self.master = master 277 """The host application.""" 278 279 self.init_master = self.default_init_master 280 self.handle_exception = self.default_handle_exception

281 282

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

284 """Handle an exception raised in the processor framework. 285 286 The function is responsible for aborting the processor by calling processor.abort() as its 287 final act. 288 289 @param processor: The processor instance. 290 @type processor: multi.processor.Processor instance 291 @param exception: The exception raised by the processor or slave processor. In the case of 292 a slave processor exception this may well be a wrapped exception of type 293 multi.processor.Capturing_exception which was raised at the point the 294 exception was received on the master processor but contains an enclosed 295 exception from a slave. 296 @type exception: Exception instance 297 """ 298 299 # Print the traceback. 300 _traceback.print_exc(file=_sys.stderr) 301 302 # Stop the processor. 303 processor.abort()

304 305

306 - def default_init_master(self, processor):

307 """Start the main loop of the host application. 308 309 @param processor: The processor instance. 310 @type processor: multi.processor.Processor instance 311 """ 312 313 self.master.run()

314 315 316

317 -class Processor_box(object):

318 """A storage class for the Processor instance and its attributes. 319 320 This singleton contains Processor instances and information about these Processors. Importantly 321 this container gives the calling code access to the Processor. 322 """ 323 324 # Class variable for storing the class instance. 325 instance = None 326

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

328 """Replacement function for implementing the singleton design pattern.""" 329 330 # First initialisation. 331 if self.instance is None: 332 self.instance = object.__new__(self, *args, **kargs) 333 334 # Already initialised, so return the instance. 335 return self.instance

336

Source Code for Package multi