{
"info": {
"author": "Alejandro Pe\u00f1a",
"author_email": "adpena@gmail.com",
"bugtrack_url": null,
"classifiers": [
"License :: OSI Approved :: MIT License",
"Operating System :: OS Independent",
"Programming Language :: Python :: 3"
],
"description": "# `achilles`\nDistributed/parallel computing in modern Python based on the `multiprocessing.Pool` API (`map`, `imap`, `imap_unordered`).\n\n## What/why is it?\nThe purpose of `achilles` is to make distributed/parallel computing as easy as possible by limiting the required configuration, hiding the details (server/node/controller architecture) and exposing a simple interface based on the popular `multiprocessing.Pool` API.\n\n> `achilles` provides developers with entry-level capabilities for concurrency across a network of machines (see PEP 372 on the intent behind adding `multiprocessing` to the standard library -> https://www.python.org/dev/peps/pep-0371/) using a server/node/controller architecture.\n\nThe `achilles_server`, `achilles_node` and `achilles_controller` are designed to run cross-platform/cross-architecture. The server/node/controller may be hosted on a single machine (for development) or deployed across heterogeneous resources.\n\n`achilles` is comparable to excellent Python packages like `pathos/pyina`, `Parallel Python` and `SCOOP`, but different in certain ways:\n- Designed for developers familiar with the `multiprocessing` module in the standard library with simplicity and ease of use in mind.\n- In addition to the blocking `map` API which requires that developers wait for all computation to be finished before accessing results (common in such packages), `imap`/`imap_unordered` allow developers to process results as they are returned to the `achilles_controller` by the `achilles_server`.\n- `achilles` allows for composable scalability and novel design patterns as:\n - Iterables including lists, lists of lists and generator functions (as first-class object - generator expressions will not work as generators cannot be serialized by `pickle`/`dill`) are accepted as arguments.\n - TIP: Use generator functions together with `imap` or `imap_unordered` to perform distributed computation on arbitrarily large data.\n - The `dill` serializer is used to transfer data between the server/node/controller and `multiprocess` (fork of `multiprocessing` that uses the `dill` serializer instead of `pickle`) is used to perform `Pool.map` on the `achilles_nodes`, so developers are freed from some of the constraints of the `pickle` serializer.\n
\n\n### Install\n`pip install achilles`\n\n### Quick Start\nStart an `achilles_server` listening for connections from `achilles_nodes` at a certain endpoint specified as arguments or in an `.env` file in the `achilles` package's directory.\n\nThen simply import `map`, `imap`, and/or `imap_unordered` from `achilles_main` and use them dynamically in your own code (under the hood they create and close `achilles_controller`s).\n\n`map`, `imap` and `imap_unordered` will distribute your function to each `achilles_node` connected to the `achilles_server`. Then, the `achilles_server` will distribute arguments to each `achilles_node` (load balanced and made into a list of arguments if the arguments' type is not already a list) which will then perform your function on the arguments using `multiprocess.Pool.map`.\n\nEach `achilles_node` finishes its work, returns the results to the `achilles_server` and waits to receive another argument. This process is repeated until all of the arguments have been exhausted.\n\n1) `runAchillesServer(host=None, port=None, username=None, secret_key=None)` -> run on your local machine or on another machine connected to your network\n\n `in:`\n ```python\n from achilles.lineReceiver.achilles_server import runAchillesServer\n\n # host = IP address of the achilles_server\n # port = port to listen on for connections from achilles_nodes (must be an int)\n # username, secret_key used for authentication with achilles_controller\n\n runAchillesServer(host='127.0.0.1', port=9999, username='foo', secret_key='bar')\n ```\n\n ```python\n\n # OR generate an .env file with a default configuration so that\n # arguments are no longer required to runAchillesServer()\n\n # use genConfig() to overwrite\n\n from achilles.lineReceiver.achilles_server import runAchillesServer, genConfig\n\n genConfig(host='127.0.0.1', port=9999, username='foo', secret_key='bar')\n runAchillesServer()\n ```\n
\n\n `out:`\n ```\n ALERT: achilles_server initiated at 127.0.0.1:9999\n Listening for connections...\n ```\n\n2) `runAchillesNode(host=None, port=None)` -> run on your local machine or on another machine connected to your network\n\n `in:`\n ```python\n from achilles.lineReceiver.achilles_node import runAchillesNode\n\n # genConfig() is also available in achilles_node, but only expects host and port arguments\n\n runAchillesNode(host='127.0.0.1', port=9999)\n ```\n
\n\n `out:`\n ```\n GREETING: Welcome! There are currently 1 open connections.\n\n Connected to achilles_server running at 127.0.0.1:9999\n CLIENT_ID: 0\n ```\n\n3) Examples of how to use the 3 most commonly used `multiprocessing.Pool` methods in `achilles`:\n
\n >> Note: `map`, `imap` and `imap_unordered` currently accept iterables including - but not limited - to lists, lists of lists, and generator functions as `achilles_args`.\n\n >> Also note: if there isn't already a `.env` configuration file in the `achilles` package directory, must use `genConfig(host, port, username, secret_key)` before using or include `host`, `port`, `username` and `secret_key` as arguments when using `map`, `imap`, `imap_unordered`.\n\n 1) `map(func, args, callback=None, chunksize=1, host=None, port=None, username=None, secret_key=None)`
\n `in:`\n ```python\n from achilles.lineReceiver.achilles_main import map\n\n def achilles_function(arg):\n return arg ** 2\n\n def achilles_callback(result):\n return result ** 2\n\n if __name__ == \"__main__\":\n results = map(achilles_function, [1, 2, 3, 4, 5, 6, 7, 8, 9, 10], achilles_callback, chunksize=1)\n print(results)\n ```\n
\n\n `out:`\n ```\n ALERT: Connection to achilles_server at 127.0.0.1:9999 and authentication successful.\n\n [[1, 16, 81, 256, 625, 1296, 2401, 4096], [6561, 10000]]\n ```\n\n 2) `imap(func, args, callback=None, chunksize=1, host=None, port=None, username=None, secret_key=None)`
\n `in:`\n ```python\n from achilles.lineReceiver.achilles_main import imap\n\n def achilles_function(arg):\n return arg ** 2 \n\n def achilles_callback(result):\n return result ** 2\n\n if __name__ == \"__main__\":\n for result in imap(achilles_function, [1,2,3,4,5,6,7,8,9,10], achilles_callback, chunksize=1):\n print(result)\n ```\n
\n\n `out:`\n ``` \n ALERT: Connection to achilles_server at 127.0.0.1:9999 and authentication successful.\n\n {'ARGS_COUNTER': 0, 'RESULT': [1, 16, 81, 256, 625, 1296, 2401, 4096]}\n {'ARGS_COUNTER': 8, 'RESULT': [6561, 10000]}\n ```\n\n 3) `imap_unordered(func, args, callback=None, chunksize=1, host=None, port=None, username=None, secret_key=None)`
\n `in:`\n ```python\n from achilles.lineReceiver.achilles_main import imap_unordered\n\n def achilles_function(arg):\n return arg ** 2\n\n def achilles_callback(result):\n return result ** 2\n\n if __name__ == \"__main__\":\n for result in imap_unordered(achilles_function, [1,2,3,4,5,6,7,8,9,10], achilles_callback, chunksize=1):\n print(result)\n ```\n
\n\n `out:`\n ``` \n ALERT: Connection to achilles_server at 127.0.0.1:9999 and authentication successful.\n\n {'ARGS_COUNTER': 8, 'RESULT': [6561, 10000]}\n {'ARGS_COUNTER': 0, 'RESULT': [1, 16, 81, 256, 625, 1296, 2401, 4096]}\n ```\n\n## How `achilles` works \n\n### Under the hood\n- `Twisted`\n - An event-driven networking engine written in Python and MIT licensed.\n- `dill`\n - `dill` extends Python\u00e2\u20ac\u2122s `pickle` module for serializing and de-serializing Python objects to the majority of the built-in Python types.\n- `multiprocess`\n - multiprocess is a fork of multiprocessing that uses `dill` instead of `pickle` for serialization. `multiprocessing` is a package for the Python language which supports the spawning of processes using the API of the standard library\u00e2\u20ac\u2122s threading module.\n\n### Examples\nSee the `examples` directory for tutorials on various use cases, including:\n- Square numbers/run multiple jobs sequentially\n- Word count (TO DO)\n\n### How to kill cluster\n```python\nfrom achilles.lineReceiver.achilles_main import killCluster\n\n# simply use the killCluster() command and verify your intent at the prompt\n# killCluster() will search for an .env configuration file in the achilles package's directory\n\n# if it does not exist, specify host, port, username and secret_key as arguments\n# a command is sent to all connected achilles_nodes to stop the Twisted reactor and exit() the process\n\n# optionally, you can pass command_verified=True to proceed directly with killing the cluster\n\nkillCluster(command_verified=True)\n```\n\n### Caveats/Things to know\n- `achilles_node`s use all of the CPU cores available on the host machine to perform `multiprocess.Pool.map` (`pool = multiprocess.Pool(multiprocess.cpu_count())`).\n- `achilles` leaves it up to the developer to ensure that the correct packages are installed on `achilles_node`s to perform the function distributed by the `achilles_server` on behalf of the `achilles_controller`. Current recommended solution is to SSH into each machine and `pip install` a `requirements.txt` file.\n- All import statements required by the developer's function, arguments and callback must be included in the definition of the function.\n- The `achilles_server` is currently designed to handle one job at a time. For more complicated projects, I highly recommend checking out `Dask` (especially `dask.distributed`) and learning more about directed acyclic graphs (DAGs).\n- Fault tolerance: if some `achilles_node` disconnects before returning expected results, the argument will be distributed to another `achilles_node` for computation instead of being lost.\n- `callback_error` argument has yet to be implemented, so detailed information regarding errors can only be gleaned from the interpreter used to launch the `achilles_server`, `achilles_node` or `achilles_controller`. Deploying the server/node/controller on a single machine is recommended for development.\n- `achilles` performs load balancing at runtime and assigns `achilles_node`s arguments by `cpu_count` * `chunksize`.\n - Default `chunksize` is 1.\n - Increasing the `chunksize` is an easy way to speed up computation and reduce the amount of time spent transferring data between the server/node/controller.\n- If your arguments are already lists, the `chunksize` argument is not used.\n - Instead, one argument/list will be distributed to the connected `achilles_node`s at a time.\n- If your arguments are load balanced, the results returned are contained in lists of length `achilles_node's cpu_count` * `chunksize`.\n - `map`:\n - Final result of `map` is an ordered list of load balanced lists (the final result is not flattened).\n - `imap`:\n - Results are returned as computation is finished in dictionaries that include the following keys:\n - `RESULT`: load balanced list of results.\n - `ARGS_COUNTER`: index of first argument (0-indexed).\n - Results are ordered.\n - The first result will correspond to the next result after the last result in the preceding results packet's list of results.\n - Likely to be slower than `immap_unordered` due to `achilles_controller` yielding ordered results. `imap_unordered` (see below) yields results as they are received, while `imap` yields results as they are received only if the argument's `ARGS_COUNTER` is expected based on the length of the `RESULT` list in the preceding results packet. Otherwise, a `result_buffer` is checked for the results packet with the expected `ARGS_COUNTER` and the current results packet is added to the `result_buffer`. If it is not found, `achilles_controller` will not yield results until a results packet with the expected `ARGS_COUNTER` is received.\n - `imap_unordered`:\n - Results are returned as computation is finished in dictionaries that include the following keys:\n - `RESULT`: load balanced list of results.\n - `ARGS_COUNTER`: index of first argument (0-indexed).\n - Results are not ordered.\n - Results packets are yielded as they are received (after any `achilles_callback` has been performed on it).\n - Fastest way of consuming results received from the `achilles_server`.\n\n\n