├── .gitignore ├── mpi4.py.sbatch ├── mpi4.py ├── README.md └── LICENSE /.gitignore: -------------------------------------------------------------------------------- 1 | #SPDX-License-Identifier: CC0-1.0 2 | 3 | # SLURM stuff 4 | slurm-* 5 | 6 | # Editor stuff 7 | *~ 8 | .*.swp 9 | 10 | # MyPy stuff 11 | .mypy_cache 12 | -------------------------------------------------------------------------------- /mpi4.py.sbatch: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # -*- coding: utf-8 -*- 3 | # vim: ts=4 sw=4 et 4 | 5 | # Copyright © 2019 The Board of Trustees of the Leland Stanford Junior University. 6 | # SPDX-License-Identifier: GPL-3.0-only 7 | 8 | # This shows how you can run mpi4.py via a SLURM batch script. 9 | # It was written by A. Karl Kornel 10 | 11 | # Set some of the job's resource requirements. 12 | # This requires that you have access to at least two nodes. 13 | # The main thing missing from here is --ntasks, the number of tasks to run. 14 | # This should be set as part of running `sbatch`. 15 | # You might also need to set things like --partition or --qos. 16 | ##SBATCH --time 0:05:00 17 | ##SBATCH --cpus-per-task 1 18 | ##SBATCH --mem-per-cpu 512M 19 | ##SBATCH --nodes 2- 20 | 21 | # This batch script will run on one of the compute nodes. 22 | # You could actually do this interactively, if you ran `srun` with all of the 23 | # SLURM options that are mentioned above (`srun --time 0:05:00 … /bin/bash`). 24 | # The environment in which this batch script runs has all of the 25 | # information (via SLURM_* environment variables) to leverage resources on 26 | # all of the nodes. It does this by calling either `srun` or `mpirun`. 27 | 28 | # First, bring things into our environment. 29 | # This batch script is set up for Stanford's FarmShare environment, 30 | # so we just bring in OpenMPI. You might need to bring in Python. 31 | # Or, maybe you don't need to load anything! 32 | module load openmpi/3.0.0 33 | 34 | # What Python do we run? Either set the PYTHON_PROGRAM environment variable, 35 | # or we will default to whatever `python3` happens to be. 36 | # We support Python 2.7, and Python 3.5 or later. 37 | PYTHON_PROGRAM=${PYTHON_PROGRAM:-python3} 38 | 39 | # What MPI runner do we use? 40 | # This is important, because it sets up all of the inter-process communications, 41 | # on both this node and all of the other nodes running this job. 42 | # We prefer letting SLURM handle MPI setup, via `srun --mpi=openmpi`. 43 | # But, you can override it by setting the RUN_COMMAND environment variable. 44 | # NOTE that SLURM's slurm.conf specifies a default MPI to use (in case `--mpi` 45 | # is not specified). FarmShare has "openmpi" set as the default, so we really 46 | # don't need to include `--mpi`. 47 | # If the default doesn't work, try one of these: 48 | # srun 49 | # srun --mpi=pmi2 50 | # srun --mpi=pmi 51 | # mpirun 52 | # NOTE that if you use mpirun, make sure your MPI has SLURM support built in. 53 | # (This is normally the case for OpenMPI.) 54 | RUN_COMMAND=${RUN_COMMAND:-srun --mpi=openmpi} 55 | 56 | # Run it! 57 | exec ${RUN_COMMAND} ${PYTHON_PROGRAM} mpi4.py 58 | -------------------------------------------------------------------------------- /mpi4.py: -------------------------------------------------------------------------------- 1 | #!/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | # vim: ts=4 sw=4 et 4 | 5 | # Copyright © 2019 The Board of Trustees of the Leland Stanford Junior University. 6 | # SPDX-License-Identifier: GPL-3.0-only 7 | 8 | """ 9 | This file shows how to use MPI in Python, using the mpi4py package! 10 | It was written by A. Karl Kornel 11 | 12 | You should run this, not with `python`, but with `mpirun`. 13 | Or, you can run it in a SLURM job, using `srun` (or `mpirun`). 14 | Read the README.md for more details! 15 | """ 16 | 17 | # Import MPI at the start, as we'll use it everywhere. 18 | from mpi4py import MPI 19 | 20 | # We'll need to query the environment 21 | from os import environ 22 | 23 | # We use type hints. Do you use type hints? You should use type hints. 24 | # NOTE: This code is compatible with Python 2.7 and 3.5+. 25 | # For Python 2.7, install `typing` from PyPI. 26 | from typing import * 27 | 28 | # Create a type variable for the types of MPI communicator objects we can use. 29 | MPIComm = Union[MPI.Intracomm, MPI.Intercomm] 30 | 31 | # Our main routine. It sanity-checks MPI and calls the correct code. 32 | def main(): 33 | # type: () -> int 34 | """Executed when called via the CLI. 35 | 36 | Performs some sanity checks, and then calls the appropriate method. 37 | """ 38 | 39 | # Get our MPI communicator, our rank, and the world size. 40 | mpi_comm = MPI.COMM_WORLD 41 | mpi_rank = mpi_comm.Get_rank() 42 | mpi_size = mpi_comm.Get_size() 43 | 44 | # Do we only have one process? If yes, then exit. 45 | if mpi_size == 1: 46 | print('You are running an MPI program with only one slot/task!') 47 | print('Are you using `mpirun` (or `srun` when in SLURM)?') 48 | print('If you are, then please use an `-n` of at least 2!') 49 | print('(Or, when in SLURM, use an `--ntasks` of at least 2.)') 50 | print('If you did all that, then your MPI setup may be bad.') 51 | return 1 52 | 53 | # Is our world size over 999? The output will be a bit weird. 54 | # NOTE: Only rank zero does anything, so we don't get duplicate output. 55 | if mpi_size >= 1000 and mpi_rank == 0: 56 | print('WARNING: Your world size {} is over 999!'.format(mpi_size)) 57 | print("The output formatting will be a little weird, but that's it.") 58 | 59 | # Sanity checks complete! 60 | 61 | # Call the appropriate function, based on our rank 62 | if mpi_rank == 0: 63 | return mpi_root(mpi_comm) 64 | else: 65 | return mpi_nonroot(mpi_comm) 66 | 67 | # This program has two parts: The controller and the worker part. 68 | # The controller is executed by rank 0; the workers by everyone else. 69 | # SOME TERMINOLOGY: 70 | # MPI World: All of the MPI processes, spawned by `mpirun` or SLURM/srun. 71 | # MPI Size: The number of MPI slots (or SLURM tasks). 72 | # Within a SLURM job, this is the number of tasks (--ntasks) for the job. 73 | # Outside of a SLURM job, this is the `-n` option given to `mpirun`. 74 | # (Or, with `mpirun`, by providing a list of hosts on which to run.) 75 | # MPI Rank: An integer number, in the range [0, MPI size). 76 | # The MPI rank is unique to each worker. 77 | 78 | # This program follows the convention that rank 0 is the "controller", and all 79 | # non-zero ranks are "workers". This is important when using things like 80 | # broadcast, or scatter/gather. But if you are only doing simple send/receive 81 | # (which we do not), then you don't need to stick to the controller-worker 82 | # paradigm. But it's still a good idea! 83 | 84 | # Our code for rank 0 85 | def mpi_root(mpi_comm): 86 | # type: (MPIComm) -> int 87 | """Routine for the controller (MPI rank 0). 88 | 89 | Generates a non-negative 32-bit number, and broadcasts it. 90 | Then, gathers responses from every other MPI rank. 91 | Each MPI rank should return a tuple of (string, int). 92 | The string is an MPI "CPU Identifier" (normally a hostname). 93 | The int is the result of `random_number + MPI_rank`. 94 | 95 | Once all results are gathered, output each result (the gathered array is 96 | sorted by MPI rank). Verify that each int returned is correct, by doing 97 | the math (`returned int == random_number + MPI_rank`) locally. 98 | 99 | At the end, send each worker (via a unicast message) an `int` zero. Then, 100 | wait for everyone to be at the same point in code (via a battier). Then 101 | we're done! 102 | """ 103 | 104 | # We import `random` here because we only use it here. 105 | import random 106 | 107 | # Get our random number, broadcast it, and print it to standard output. 108 | # NOTE: The lower-case methods (like `bcast()`) take Python object, and do 109 | # the serialization for us (yay!). 110 | # `bast()` is blocking, in the sense that it does not return until 111 | # the data have been sent, but it is _not_ synchronizing. 112 | # There's also no guarantee as to _how_ the data were conveyed. 113 | # NOTE: In Python 3.6+, we should use `secret` instead of `random`. 114 | random_number = random.randrange(2**32) 115 | mpi_comm.bcast(random_number) 116 | print('Controller @ MPI Rank 0: Input {}'.format(random_number)) 117 | 118 | # Gather all of our responses 119 | # `gather()` takes a parameter, which for the root is `None`. 120 | # Again, from an MPI perspective this is blocking, and it is synchronizing 121 | # in the sense that we know all of workers have sent something. But this 122 | # is still not proper synchronization. 123 | # NOTE: The returned array puts our parameter into slot 0. So, the length 124 | # of the returned array should match the MPI world size. 125 | GatherResponseType = List[Tuple[str, int]] 126 | response_array = mpi_comm.gather(None) # type: GatherResponseType 127 | 128 | # Sanity check: Did we get back a response from everyone? 129 | mpi_size = mpi_comm.Get_size() 130 | if len(response_array) != mpi_size: 131 | print('ERROR! The MPI world has {} members, but we only gathered {}!' 132 | .format(mpi_size, len(response_array)) 133 | ) 134 | return 1 135 | 136 | # Output each worker's results. 137 | # NOTE: We skip entry zero, because rank 0 is us, and we gathered `None`. 138 | for i in range(1, mpi_size): 139 | # Sanity check: Did we get a tuple of (string, int)? 140 | if len(response_array[i]) != 2: 141 | print('WARNING! MPI rank {} sent a mis-sized ({}) tuple!' 142 | .format( 143 | i, 144 | len(response_array[i]) 145 | ) 146 | ) 147 | continue 148 | if type(response_array[i][0]) is not str: 149 | print('WARNING! MPI rank {} sent a tuple with a {} instead of a str!' 150 | .format( 151 | i, 152 | str(type(response_array[i][0])) 153 | ) 154 | ) 155 | continue 156 | if type(response_array[i][1]) is not int: 157 | print('WARNING! MPI rank {} sent a tuple with a {} instead of an int!' 158 | .format( 159 | i, 160 | str(type(response_array[i][1])) 161 | ) 162 | ) 163 | continue 164 | 165 | # Is the result OK? Check if random_number + i == response 166 | if random_number + i == response_array[i][1]: 167 | result = 'OK' 168 | else: 169 | result = 'BAD' 170 | 171 | # Output the message. 172 | # The first field `{: >3}` translates as... 173 | # `: -> Take the next parameter from `format()` 174 | # ` ` -> Use spaces as the fill character 175 | # `>` -> Right-align 176 | # `3` -> Normal width is three characters 177 | print(' Worker at MPI Rank {: >3}: Output {} is {} (from {})' 178 | .format( 179 | i, 180 | response_array[i][1], 181 | result, 182 | response_array[i][0], 183 | ) 184 | ) 185 | 186 | # Send the worker a unidirectional MPI message, signifying 187 | # "You can exit now!". But really, this is just showing off MPI 188 | # unidirectional messaging. 189 | # NOTE: This will be slow! Since we have to (a) reach out to that 190 | # specific node, and (b) wait for that node to be ready to receive. 191 | mpi_comm.send( 192 | obj=0, 193 | dest=i, 194 | tag=0, 195 | ) 196 | 197 | # Before we finish, do an MPI synchronization barrier. 198 | # This the only proper way of doing synrhconization with MPI. 199 | # Do we need it here? Nope! We're just showing it off. 200 | mpi_comm.barrier() 201 | 202 | # We're all done! 203 | return 0 204 | 205 | # Our code for ranks 1 and up 206 | def mpi_nonroot(mpi_comm): 207 | # type: (MPIComm) -> int 208 | """Routine for the MPI workers (ranks 1 and up). 209 | 210 | Receives a number from a broadcast. 211 | To that received number, add our MPI rank (a positive, non-zero integer). 212 | 213 | Return, via the gather process, a tuple with two items: 214 | * The MPI "CPU Identifier" (normally a hostname) 215 | * The calculated number, above. 216 | 217 | Then, enter a loop: We receive a number (an `int`) from the controller. If 218 | the number is zero, we exit the loop. Otherwise, we divide the number by 219 | two, convert the result to an int, and send the result to the controller. 220 | 221 | Finally, after the loop is over, we synchronize via an MPI barrier. 222 | """ 223 | 224 | # Get our MPI rank. 225 | # This is a unique number, in the range [0, MPI_size), which identifies us 226 | # in this MPI world. 227 | mpi_rank = mpi_comm.Get_rank() 228 | 229 | # Get the number broadcast to us. 230 | # `bcast()` takes a parameter, but since we're not sending, we use `None`. 231 | # NOTE: This blocks until rank zero broadcasts something, but it's not 232 | # synchronizing, in that rank zero might have moved on already. 233 | # And again, no way to know exactly how this got to us. 234 | random_number = mpi_comm.bcast(None) # type: int 235 | 236 | # Sanity check: Did we actually get an int? 237 | if type(random_number) is not int: 238 | print('ERROR in MPI rank {}: Received a non-integer "{}" from the broadcast!' 239 | .format( 240 | mpi_rank, 241 | random_number, 242 | ) 243 | ) 244 | return 1 245 | 246 | # Our response is the random number + our rank 247 | response_number = random_number + mpi_rank 248 | 249 | # Return our response 250 | # `gather()` knows that we're not the root, and so we're returning. 251 | response = ( 252 | MPI.Get_processor_name(), 253 | response_number, 254 | ) 255 | mpi_comm.gather(response) 256 | 257 | # Receive a unidirectional message (an `int`) from the controller. 258 | # Every time we get a non-zero integer, we return `floor(int/2)`. 259 | # When we get a zero, we stop. 260 | 261 | # Move message-receiving into a helper. 262 | def get_message(mpi_comm): 263 | # type: (MPIComm) -> Union[int, None] 264 | message = mpi_comm.recv( 265 | source=0, 266 | tag=0, 267 | ) # type: int 268 | if type(message) is not int: 269 | print('ERROR in MPI rank {}: Received a non-integer message!' 270 | .format( 271 | mpi_rank, 272 | ) 273 | ) 274 | return None 275 | else: 276 | return message 277 | 278 | # Start looping! 279 | message = get_message(mpi_comm) 280 | while (message is not None) and (message != 0): 281 | # Divide the number by 2, and send it back 282 | mpi_comm.send( 283 | obj=int(message/2), 284 | dest=0, 285 | tag=0, 286 | ) 287 | 288 | # Get a new message 289 | message = get_message(mpi_comm) 290 | 291 | # Did we get an error? 292 | if message is None: 293 | return 1 294 | 295 | # Before we finish, do an MPI synchronization barrier. 296 | # This the only proper way of doing synrhconization with MPI. 297 | # Do we need it here? Nope! We're just showing it off. 298 | mpi_comm.barrier() 299 | 300 | # That's it! 301 | return 0 302 | 303 | # Run main() 304 | if __name__ == '__main__': 305 | import sys 306 | sys.exit(main()) 307 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | MPI is a complicated beast, conjuring thoughts of grey-beards sitting in front 2 | of amber screens, typing Fortran code that runs in distant, dark computing 3 | halls. 4 | 5 | But it's not like that! MPI is something that anyone can use! If you have 6 | code that would normally `fork()` (or `clone()`) itself, but your single 7 | machine does not have enough CPU cores to meet your needs, MPI is a way to run 8 | the same code across multiple machines, without needing to think about setting 9 | up or coordinating IPC. 10 | 11 | The code and docs in this repository are targeted at Python developers, and 12 | will take you from zero to running a simple MPI Python script! 13 | 14 | # Quick Start 15 | 16 | Start with an Ubuntu Bionic (18.04) system, and run these commands to install 17 | software and run this code! 18 | 19 | ``` 20 | git clone https://github.com/akkornel/mpi4py.git 21 | sudo -s 22 | apt install python3-mpi4py 23 | echo 'btl_base_warn_component_unused = 0' > /etc/openmpi/openmpi-mca-params.conf 24 | exit 25 | cd mpi4py 26 | mpirun -n 20 python3 mpi4.py 27 | ``` 28 | 29 | See it in action: 30 | [![asciicast](https://asciinema.org/a/gd64eQj99Tnvew9q27Ia9ciXe.svg)](https://asciinema.org/a/gd64eQj99Tnvew9q27Ia9ciXe?autoplay=1) 31 | 32 | The Python code is simple: It has one 'controller', and one or more 'workers': 33 | 34 | 1. The controller generates a random number, and broadcasts it to everyone. 35 | 36 | 2. Each worker has a unique ID (its "rank"), and the hostname of the machine 37 | where it is running. The worker receives the random number, adds its rank, 38 | and sends the result (the new number) and its hostname back to the 39 | controller. 40 | 41 | 3. The controller "gathers" all of the results (which come to it sorted by 42 | rank), and displays them. The output includes an "OK" if the math (random 43 | number + rank) is correct. 44 | 45 | The Python code is simple: it does not have to deal with launching workers, and 46 | it does not have to set up communications between them (inter-process 47 | communications, or IPC). 48 | 49 | The above example showed 20 copies of mpi4.py, all running on the local system. 50 | But, if you have two identical machines, you can spread the workers out across 51 | all hosts, without any changes: 52 | 53 | [![asciicast](https://asciinema.org/a/K05Tqig8AcbFv68O1qlSDFwbd.svg)](https://asciinema.org/a/K05Tqig8AcbFv68O1qlSDFwbd?autoplay=1) 54 | 55 | If you have access to a compute environment running SLURM, you can let SLURM do 56 | all of the work for you: 57 | 58 | [![asciicast](https://asciinema.org/a/XyPGpX7gTG5ZrEWWQ4KS3OzEl.svg)](https://asciinema.org/a/XyPGpX7gTG5ZrEWWQ4KS3OzEl?autoplay=1) 59 | 60 | Read on for information on the technologies that make all this work, and how to 61 | try it for yourself! 62 | 63 | # Behind the Scenes 64 | 65 | Behind the above example, there are many moving pieces. Here is an explanation 66 | of each one. 67 | 68 | ## MPI 69 | 70 | MPI is an API specification for inter-process communication. The specification 71 | defines bindings for both C/C++ and Fortran, which are used as the basis for 72 | other languages (in our case, Python). MPI provides for many IPC needs, 73 | including… 74 | 75 | * Bi-directional communication, both blocking and non-blocking 76 | 77 | * Scatter-gather and broadcast communication 78 | 79 | * Shared access to files, and to pools ("windows") of memory 80 | 81 | The latest published version of the standard is 82 | [MPI-3.1](https://www.mpi-forum.org/docs/mpi-3.1/mpi31-report.pdf), published 83 | June 4, 2015. MPI specification versions have an "MPI-" prefix to distinguish 84 | them from implementation versions. Specifications are developed by the [MPI 85 | Forum](https://www.mpi-forum.org). 86 | 87 | MPI is interesting in that it only specifies an API; it is up to other groups 88 | to provide the implementations of the API. In fact, MPI does not specify 89 | things like encodings, or wire formats, or even transports. The implementation 90 | has full reign to use (what it feels is) the most efficient encoding, and to 91 | support whatever transport it wants. Most MPI implementations support 92 | communications over Infiniband (using native Infiniband verbs) or TCP 93 | (typically over Ethernet). Newer implementations also support native 94 | communication over modern 'converged' high-performance platforms, like 95 | [RoCE](http://www.mellanox.com/related-docs/whitepapers/WP_RoCE_vs_iWARP.pdf). 96 | 97 | Because each MPI implementation has control over how the MPI API is 98 | implemented, a given "world" of programs (or, more accurately, a world of 99 | copies of a single program) must all use the same MPI implementation. 100 | 101 | ## OpenMPI 102 | 103 | [OpenMPI](https://www.open-mpi.org/) is an implementation of the MPI API, 104 | providing libraries and headers for C, C++, Fortran, and Java; and supporting 105 | interfaces from TCP to Infiniband to RoCE. OpenMPI is fully 106 | [open-source](https://github.com/open-mpi/ompi), with contributors from 107 | adademic and research fields (IU, UT, HPCC) to commercial entities (Cisco, 108 | Mellanox, and nVidia). It is licensed under the 3-clause BSD license. Your 109 | Linux distribution probably has it packaged already. 110 | 111 | **NOTE**: When running OpenMPI for the first time, you may get a warning which 112 | says "A high-performance Open MPI point-to-point messaging module was unable to 113 | find any relevant network interfaces". That is OpenMPI complaining that it is 114 | not able to find anything other than a basic Ethernet interface. 115 | 116 | To suppress the warning about no high-performance network interfaces, edit the 117 | file at path `/etc/openmpi/openmpi-mca-params.conf`, and insert the line 118 | `btl_base_warn_component_unused = 0`. 119 | 120 | ## Python 121 | 122 | As this is a Python demonstration, you will need Python! MPI has been around 123 | for a long time, and so many versions of Python are supported. The code in 124 | this repo works with Python 2.7, and also Python 3.5 and later. 125 | 126 | The Python code in this repository includes type annotations. To run in Python 127 | 2.7, you will need to install the 128 | [typing](https://pypi.org/project/typing/) package (it comes as part of the 129 | standard library in 3.5 and later). 130 | 131 | ## mpi4py 132 | 133 | There are no MPI implementations written specifically for Python. Instead, the 134 | [mpi4py](https://pypi.org/project/mpi4py/) package provides a Python wrapper 135 | around your choice of MPI implementation. In addition, mpi4py provides two 136 | ways to interface with MPI: You can provide binary data (as a `bytes` string or 137 | a `bytearray`); or you can provide a Python object, which mpi4py will serialize 138 | for you (using `pickle`). This adds a massive convenience for Python 139 | programmers, because you do not have to deal with serialization (other than 140 | ensuring your Python objects are pickleable). 141 | 142 | **WARNING**: Although mpi4py will handle serialization for you, you are still 143 | responsible for checking what data are being received. It is your code running 144 | at the other end, but you may not be sure exactly what state the other copy is 145 | in right now. 146 | 147 | **NOTE**: If you plan on building mpi4py yourself, you should be aware that MPI 148 | was created at a time when 149 | [`pkg-config`](https://www.freedesktop.org/wiki/Software/pkg-config/) did not 150 | exist. Instead of storing things like compiler flags in a `.pc` file, it 151 | provides a set of wrapper scripts (`mpicc`, and the like), which you should 152 | call instead of calling the normal compiler programs. The commands are built 153 | as part of a normal OpenMPI installation, and mpi4py's `setup.py` script knows 154 | to use them, and so `pip install --user mpi4py` should work. 155 | 156 | ## Tested Configurations 157 | 158 | The code in this repository was tested by the author on the following 159 | OS, OpenMPI, and mpi4py configurations, on an x64 platform: 160 | 161 | * Ubuntu 16.04.6 162 | 163 | OpenMPI 3.0.0 (built from upstream) was used in all tests. 164 | 165 | mpi4py version 3.0.2 (built from upstream) was used in all tests. The 166 | Ubuntu-supplied version (1.10.2-8ubuntu1) was tested, but reliably crashed 167 | with a buffer overflow. 168 | 169 | The above software was tested in the following Python versions: 170 | 171 | * Python 2.7 (2.7.12-1~16.04) 172 | 173 | * Python 3.5 (3.5.1-3) 174 | 175 | * Python 3.6.8 from upstream 176 | 177 | * Python 3.7.3 from upstream 178 | 179 | * Ubuntu 18.04.2 180 | 181 | OpenMPI 2.1.1 (2.1.1-8) was used in all tests. 182 | 183 | mpi4py 2.0.0 (2.0.0-3) was used in all tests. 184 | 185 | The above software was tested in the following Python versions: 186 | 187 | * Python 2.7 (2.7.15~rc1-1). 188 | 189 | * Python 3.6 (3.6.7-1~18.04). 190 | 191 | # Running 192 | 193 | You can run this demonstration code multiple ways. This documentation explains 194 | how to run directly (with `mpirun`). This works automatically on one machine, 195 | and can be configured to run on multiple machines. Information is also 196 | provided for people with access to a SLURM environment. 197 | 198 | ## mpirun 199 | 200 | The easiest way to run your code is with the `mpirun` command. This command, 201 | run in a shell, will launch multiple copies of your code, and set up 202 | communications between them. 203 | 204 | Here is an example of running `mpirun` on a single machine, to launch five 205 | copies of the script: 206 | 207 | ``` 208 | akkornel@machine1:~/mpi4py$ mpirun -n 5 python3 mpi4.py 209 | Controller @ MPI Rank 0: Input 773507788 210 | Worker at MPI Rank 1: Output 773507789 is OK (from machine1) 211 | Worker at MPI Rank 2: Output 773507790 is OK (from machine1) 212 | Worker at MPI Rank 3: Output 773507791 is OK (from machine1) 213 | Worker at MPI Rank 4: Output 773507792 is OK (from machine1) 214 | ``` 215 | 216 | The `-n` (or `-np`) option specifies the number of MPI tasks to run. 217 | 218 | Note above how all of the MPI tasks ran on the same machine. If you want to 219 | run your code across multiple machines (which is probably the reason why you 220 | came here!), there are some steps you'll need to take: 221 | 222 | * All machines should be running the same architecture, and the same versions 223 | of OS, MPI implementation, Python, and mpi4py. 224 | 225 | Minor variations in the version numbers may be tolerable (for example, one 226 | machine running Ubuntu 18.04.1 and one machine running Ubuntu 18.04.2), but 227 | you should examine what is different between versions. For example, if one 228 | machine is running a newer mpi4py, and that version fixed a problem that 229 | applies to your code, you should make sure mpi4py is *at least* that version 230 | across all machines. 231 | 232 | * There should be no firewall block: Each host needs to be able to communicate 233 | with each other host, on any TCP or UDP port. You can limit the ports used, 234 | but that is out of the scope of this quick-start. 235 | 236 | * SSH should be configured so that you can get an SSH prompt on each machine 237 | without needing to enter a password or answer a prompt. In other words, 238 | you'll need public-key auth configured. 239 | 240 | * The code should either live on shared storage, or should be identical on all 241 | nodes. 242 | 243 | `mpirun` will not copy your code to each machine, so it needs to be present 244 | at the same path on all nodes. If you do not have shared storage available, 245 | one way to ensure this is to keep all of your code in a version-controlled 246 | repository, and check the commit ID (or revision number, etc.) before 247 | running. 248 | 249 | Assuming all of those conditions are met, you just need to tell `mpirun` what 250 | hosts to use, like so: 251 | 252 | ``` 253 | akkornel@machine1:~/mpi4py$ cat host_list 254 | machine1 255 | machine2 256 | akkornel@machine1:~/mpi4py$ mpirun -n 5 --hostfile host_list python3 mpi4.py 257 | Controller @ MPI Rank 0: Input 1421091723 258 | Worker at MPI Rank 1: Output 1421091724 is OK (from machine1) 259 | Worker at MPI Rank 2: Output 1421091725 is OK (from machine1) 260 | Worker at MPI Rank 3: Output 1421091726 is OK (from machine2) 261 | Worker at MPI Rank 4: Output 1421091727 is OK (from machine2) 262 | ``` 263 | 264 | OpenMPI will spread out the MPI tasks evenly across nodes. If you would like 265 | to limit how many tasks a host may run, add `slots=X` to the file, like so: 266 | 267 | ``` 268 | machine1 slots=2 269 | machine2 slots=4 270 | ``` 271 | 272 | In the above example, machine1 has two cores (or, one hyper-threaded core) and 273 | machine2 has four cores. 274 | 275 | **NOTE**: If you specify a slot count on *any* machine, `mpirun` will fail if 276 | you ask for more slots that what is available. 277 | 278 | And that's it! You should now be up and running with MPI across multiple 279 | hosts. 280 | 281 | ## SLURM 282 | 283 | Many compute environments, especially in HPC, use the 284 | [SLURM](https://slurm.schedmd.com) job scheduler. SLURM is able to communicate 285 | node and slot information to programs that use MPI, and in some MPI 286 | implementations SLURM is able to do all of the communications setup on its own. 287 | 288 | With SLURM, there are two ways of launching our MPI job. The first is to use 289 | `srun`, launching the job in a synchronous fasion (that was shown in the 290 | example at the top of this page). The second is to use `sbatch`, providing a 291 | batch script that will be run asynchronously. 292 | 293 | Regardless of the method of execution (that is, regardless of the SLURM command 294 | we run), we will need to tell SLURM what resources we will require: 295 | 296 | * **Time**. Our job takes less than five minutes to complete. 297 | 298 | * **Number of Tasks**. This is the equivalent to the `-n` option for `mpirun`. 299 | We need to specify how many MPI tasks to run. 300 | 301 | * **CPUs/cores**. Our code is single-threaded, so this is "1 per task". 302 | 303 | * **Memory**. We are running a single Python interpreter in each task, but 304 | let's overestimate and say "512 MB per core". Note that SLURM specifies 305 | memory either per core or per node, not per task. 306 | 307 | Note that in our request we never specified how many machines to run on. As 308 | long as all machines have access to the same storage (which is normal in a 309 | compute environment), we do not care exactly where our program runs. All of 310 | our tasks may run on a single machine, or they may be spread out across 311 | multiple machines. 312 | 313 | Note also that SLURM has support for many different MPI implementations, but in 314 | some cases, special care may need to be taken. For OpenMPI, nothing special is 315 | required: SLURM supports OpenMPI natively, and OpenMPI supports SLURM by 316 | default. More information on other MPI implementations is available [from 317 | SLURM](https://slurm.schedmd.com/mpi_guide.html). To see which MPIs are 318 | supported by your SLURM environment, run the command `srun --mpi list`. You 319 | can check the default by looking at the environment's `slurm.conf` file (which 320 | is normally located in `/etc` on each machine). The setting is named 321 | "MpiDefault". 322 | 323 | In the examples which follow, we will always be requesting at least two nodes. 324 | 325 | ## Synchronous SLURM with salloc 326 | 327 | Running "synchronously" means you submit the work to SLURM, and you wait for 328 | results to come back. Any messages output by your code (for example, via 329 | standard output or standard error) will be displayed in the terminal. 330 | 331 | To run synchronously, we use the `srun` command, like so: 332 | 333 | ``` 334 | akkornel@rice15:~/mpi4py$ module load openmpi/3.0.0 335 | akkornel@rice15:~/mpi4py$ srun --time 0:05:00 --ntasks 10 --cpus-per-task 1 --mem-per-cpu 512M --nodes 2-10 python3 mpi4.py 336 | Controller @ MPI Rank 0: Input 3609809428 337 | Worker at MPI Rank 1: Output 3609809429 is OK (from wheat09) 338 | Worker at MPI Rank 2: Output 3609809430 is OK (from wheat10) 339 | Worker at MPI Rank 3: Output 3609809431 is OK (from wheat11) 340 | Worker at MPI Rank 4: Output 3609809432 is OK (from wheat12) 341 | Worker at MPI Rank 5: Output 3609809433 is OK (from wheat13) 342 | Worker at MPI Rank 6: Output 3609809434 is OK (from wheat14) 343 | Worker at MPI Rank 7: Output 3609809435 is OK (from wheat15) 344 | Worker at MPI Rank 8: Output 3609809436 is OK (from wheat16) 345 | Worker at MPI Rank 9: Output 3609809437 is OK (from wheat17) 346 | ``` 347 | 348 | In the above example, OpenMPI is provided via an [Lmod 349 | module](https://lmod.readthedocs.io/en/latest/), instead of a system package. 350 | We have already built mpi4py using that version of OpenMPI. So, before running 351 | our code, we first load the same OpenMPI module used to build mpi4py. 352 | 353 | (We don't need to run `module load` on the remote system, because Lmod modules 354 | do their work by manipulating environment variables, which SLURM copies over to 355 | the compute nodes.) 356 | 357 | Once that is complete, we use the `srun` command to run our job. SLURM finds 358 | enough nodes to run our job, runs it, and prints the contents of standard 359 | output and standard error. 360 | 361 | Note that the above example assumes OpenMPI is configured as SLURM's default 362 | MPI. If not , you can provide `--mpi=openmpi` to the `srun` command, to force 363 | it to use native OpenMPI. 364 | 365 | ## SLURM batch jobs with sbatch 366 | 367 | When running as a batch job, most of the resource options should be specified 368 | in the script that is provided to `sbatch`. That script should include (via 369 | `##SBATCH` special comments) all of the resource-setting options that you 370 | previously included in the `srun` command line, *except* for `--ntasks` (and 371 | possibly also `--time`). 372 | 373 | Since the code supports any number of workers (as is typical for an MPI 374 | program), you should provide `--ntasks` on the `sbatch` command-line, so that 375 | you can adjust it to fit your current needs. 376 | 377 | This example uses the sbatch script included in this repository. Ten workers 378 | are requested. 379 | 380 | ``` 381 | akkornel@rice15:~/mpi4py$ sbatch --ntasks 10 mpi4.py.sbatch 382 | Submitted batch job 1491736 383 | akkornel@rice15:~/mpi4py$ sleep 10 384 | akkornel@rice15:~/mpi4py$ squeue -j 1491736 385 | JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) 386 | akkornel@rice15:~/mpi4py$ cat slurm-1491736.out 387 | Controller @ MPI Rank 0: Input 3761677288 388 | Worker at MPI Rank 1: Output 3761677289 is OK (from wheat08) 389 | Worker at MPI Rank 2: Output 3761677290 is OK (from wheat08) 390 | Worker at MPI Rank 3: Output 3761677291 is OK (from wheat08) 391 | Worker at MPI Rank 4: Output 3761677292 is OK (from wheat08) 392 | Worker at MPI Rank 5: Output 3761677293 is OK (from wheat08) 393 | Worker at MPI Rank 6: Output 3761677294 is OK (from wheat08) 394 | Worker at MPI Rank 7: Output 3761677295 is OK (from wheat08) 395 | Worker at MPI Rank 8: Output 3761677296 is OK (from wheat08) 396 | Worker at MPI Rank 9: Output 3761677297 is OK (from wheat08) 397 | ``` 398 | 399 | When SLURM launches the job, the batch script is executed on one of the nodes 400 | allocated to the job. The batch script is responsible for setting up the 401 | environment (in this case, loading the OpenMPI module), and then executing 402 | `srun`. Note that in the batch script, `srun` does not have any command-line 403 | options (except for `--mpi`); `srun` is able to recognize that it is running 404 | inside a batch job, and will automatically scale itself to use all of the 405 | resources available. 406 | 407 | That's it! Without having to change any Python code, your work is now being 408 | run via a job scheduler, with all of the resource scheduling and IPC facilities 409 | manged for you. 410 | 411 | # Where to go next? 412 | 413 | This was only a dip of the toe into the world of MPI. To continue on, you 414 | should first access the [mpi4py 415 | documentation](https://mpi4py.readthedocs.io/en/stable/). As most MPI 416 | publications provide their code in C, the mpi4py docs will guide you on how to 417 | convert that C code into Python calls. You can also use the built-in help 418 | system, like so: 419 | 420 | ``` 421 | akkornel@rice13:~$ python3.5 422 | Python 3.5.2 (default, Nov 12 2018, 13:43:14) 423 | [GCC 5.4.0 20160609] on linux 424 | Type "help", "copyright", "credits" or "license" for more information. 425 | >>> from mpi4py import MPI 426 | >>> mpi_comm = MPI.COMM_WORLD 427 | >>> help(mpi_comm.Get_rank) 428 | Help on built-in function Get_rank: 429 | 430 | Get_rank(...) method of mpi4py.MPI.Intracomm instance 431 | Comm.Get_rank(self) 432 | 433 | Return the rank of this process in a communicator 434 | ``` 435 | 436 | The mpi4py docs also has a tutorial, which you should read! 437 | 438 | Now that is done, the next thing to do is to think of a problem, and to try 439 | solving it using MPI! Focus on a problem that can be parallelized easily. 440 | 441 | The Lawrence Livermore National Lab has a tutorial on MPI, featuring some 442 | exercises that you can try. For example, try [going to the 443 | tutorial](https://computing.llnl.gov/tutorials/mpi/) and skipped ahead to the 444 | _Point to Point Communication Routines_ section, which guides you through one 445 | of the ways of calculating the digits of Pi. 446 | 447 | Other tutorials include Rajeev Thakur's [Introduction to 448 | MPI](https://courses.cs.washington.edu/courses/csep524/13wi/Thakur-MPI.pdf) 449 | slide deck, which has been condenced by Brad Chamberlain of the University of 450 | Washington (for a course there); and the [MPI for 451 | Dummies](https://htor.inf.ethz.ch/teaching/mpi_tutorials/ppopp13/2013-02-24-ppopp-mpi-basic.pdf) 452 | slide deck, created for the [PPoPP 453 | conference](https://conf.researchr.org/series/PPoPP). 454 | 455 | If you are interested in spreading your MPI code out across multiple machines, 456 | you might want to set up a SLURM environment. First, get some shared storage 457 | by setting up an NFS server (in 458 | [Ubuntu](https://help.ubuntu.com/community/SettingUpNFSHowTo), or you could use 459 | [Amazon EFS](https://aws.amazon.com/efs/) or [Google Cloud 460 | Filestore](https://cloud.google.com/filestore/)). Then, check out the [SLURM 461 | Quick Start](https://slurm.schedmd.com/quickstart_admin.html). On RPM-based 462 | systems, the quick-start walks you through creating RPMs; on Debian-based 463 | systems, the packages are already available under the name "slurm-llnl". 464 | 465 | Once you are done with all that, start looking at nearby universities and 466 | research institutions. They may have a job opening for you! 467 | 468 | # Copyright & License 469 | 470 | The contents of this repository are Copyright © 2019 the Board of Trustees of 471 | the Leland Stanford Junior University. 472 | 473 | The code and docs (with the exception of the `.gitignore` file) are licensed 474 | under the GNU General Public License, version 3.0. 475 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | --------------------------------------------------------------------------------