achilles

achilles

Distributed/parallel computing in modern Python based on the multiprocessing.Pool API (map, imap, imap_unordered).

What/why is it?

The 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.

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.

The 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.

achilles is comparable to excellent Python packages like pathos/pyina, Parallel Python and SCOOP, but different in certain ways:

• Designed for developers familiar with the multiprocessing module in the standard library with simplicity and ease of use in mind.
• 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.
• achilles allows for composable scalability and novel design patterns as:
  • 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.
    • TIP: Use generator functions together with imap or imap_unordered to perform distributed computation on arbitrarily large data.
  • 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.
    

Install

pip install achilles

Quick Start

Start 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.

Then 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_controllers).

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.

Each 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.

1. runAchillesServer(host=None, port=None, username=None, secret_key=None) -> run on your local machine or on another machine connected to your network

   in:

   from achilles.lineReceiver.achilles_server import runAchillesServer
   
   # host = IP address of the achilles_server
   # port = port to listen on for connections from achilles_nodes (must be an int)
   # username, secret_key used for authentication with achilles_controller
   
   runAchillesServer(host='127.0.0.1', port=9999, username='foo', secret_key='bar')
   

   
   # OR generate an .env file with a default configuration so that
   # arguments are no longer required to runAchillesServer()
   
   # use genConfig() to overwrite
   
   from achilles.lineReceiver.achilles_server import runAchillesServer, genConfig
   
   genConfig(host='127.0.0.1', port=9999, username='foo', secret_key='bar')
   runAchillesServer()
   

   
   

   out:

   ALERT: achilles_server initiated at 127.0.0.1:9999
   Listening for connections...
   

2. runAchillesNode(host=None, port=None) -> run on your local machine or on another machine connected to your network

   in:

   from achilles.lineReceiver.achilles_node import runAchillesNode
   
   # genConfig() is also available in achilles_node, but only expects host and port arguments
   
   runAchillesNode(host='127.0.0.1', port=9999)
   

   
   

   out:

   GREETING: Welcome! There are currently 1 open connections.
   
   Connected to achilles_server running at 127.0.0.1:9999
   CLIENT_ID: 0
   

3. Examples of how to use the 3 most commonly used multiprocessing.Pool methods in achilles:
   

   Note: map, imap and imap_unordered currently accept iterables including - but not limited - to lists, lists of lists, and generator functions as achilles_args.

   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.

   1. map(func, args, callback=None, chunksize=1, host=None, port=None, username=None, secret_key=None)

in:

      from achilles.lineReceiver.achilles_main import map
      
      def achilles_function(arg):
         return arg ** 2
      
      def achilles_callback(result):
         return result ** 2
      
      if __name__ == "__main__":
         results = map(achilles_function, [1, 2, 3, 4, 5, 6, 7, 8, 9, 10], achilles_callback, chunksize=1)
         print(results)
      

      
      

      out:

      ALERT: Connection to achilles_server at 127.0.0.1:9999 and authentication successful.
      
      [[1, 16, 81, 256, 625, 1296, 2401, 4096], [6561, 10000]]
      

   2. imap(func, args, callback=None, chunksize=1, host=None, port=None, username=None, secret_key=None)

in:

      from achilles.lineReceiver.achilles_main import imap
      
      def achilles_function(arg):
        return arg ** 2 
      
      def achilles_callback(result):
        return result ** 2
      
      if __name__ == "__main__":
        for result in imap(achilles_function, [1,2,3,4,5,6,7,8,9,10], achilles_callback, chunksize=1):
            print(result)
      

      
      

      out:

      ALERT: Connection to achilles_server at 127.0.0.1:9999 and authentication successful.
      
      {'ARGS_COUNTER': 0, 'RESULT': [1, 16, 81, 256, 625, 1296, 2401, 4096]}
      {'ARGS_COUNTER': 8, 'RESULT': [6561, 10000]}
      

   3. imap_unordered(func, args, callback=None, chunksize=1, host=None, port=None, username=None, secret_key=None)

in:

      from achilles.lineReceiver.achilles_main import imap_unordered
      
      def achilles_function(arg):
          return arg ** 2
      
      def achilles_callback(result):
          return result ** 2
      
      if __name__ == "__main__":
          for result in imap_unordered(achilles_function, [1,2,3,4,5,6,7,8,9,10], achilles_callback, chunksize=1):
              print(result)
      

      
      

      out:

      ALERT: Connection to achilles_server at 127.0.0.1:9999 and authentication successful.
      
      {'ARGS_COUNTER': 8, 'RESULT': [6561, 10000]}
      {'ARGS_COUNTER': 0, 'RESULT': [1, 16, 81, 256, 625, 1296, 2401, 4096]}
      

How achilles works

Under the hood

• Twisted
  • An event-driven networking engine written in Python and MIT licensed.
• dill
  • dill extends Python’s pickle module for serializing and de-serializing Python objects to the majority of the built-in Python types.
• multiprocess
  • 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’s threading module.

Examples

See the examples directory for tutorials on various use cases, including:

• Square numbers/run multiple jobs sequentially
• Word count (TO DO)

How to kill cluster

from achilles.lineReceiver.achilles_main import killCluster

# simply use the killCluster() command and verify your intent at the prompt
# killCluster() will search for an .env configuration file in the achilles package's directory

# if it does not exist, specify host, port, username and secret_key as arguments
# a command is sent to all connected achilles_nodes to stop the Twisted reactor and exit() the process

# optionally, you can pass command_verified=True to proceed directly with killing the cluster

killCluster(command_verified=True)

Caveats/Things to know

• achilles_nodes use all of the CPU cores available on the host machine to perform multiprocess.Pool.map (pool = multiprocess.Pool(multiprocess.cpu_count())).
• achilles leaves it up to the developer to ensure that the correct packages are installed on achilles_nodes 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.
• All import statements required by the developer's function, arguments and callback must be included in the definition of the function.
• 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).
• 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.
• 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.
• achilles performs load balancing at runtime and assigns achilles_nodes arguments by cpu_count * chunksize.
  • Default chunksize is 1.
  • 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.
• If your arguments are already lists, the chunksize argument is not used.
  • Instead, one argument/list will be distributed to the connected achilles_nodes at a time.
• If your arguments are load balanced, the results returned are contained in lists of length achilles_node's cpu_count * chunksize.
  • map:
    • Final result of map is an ordered list of load balanced lists (the final result is not flattened).
  • imap:
    • Results are returned as computation is finished in dictionaries that include the following keys:
      • RESULT: load balanced list of results.
      • ARGS_COUNTER: index of first argument (0-indexed).
    • Results are ordered.
      • The first result will correspond to the next result after the last result in the preceding results packet's list of results.
      • 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.
  • imap_unordered:
    • Results are returned as computation is finished in dictionaries that include the following keys:
      • RESULT: load balanced list of results.
      • ARGS_COUNTER: index of first argument (0-indexed).
    • Results are not ordered.
      • Results packets are yielded as they are received (after any achilles_callback has been performed on it).
      • Fastest way of consuming results received from the achilles_server.

---

achilles is in the early stages of active development and your suggestions/contributions are kindly welcomed.

achilles is written and maintained by Alejandro Peña. Email me at adpena at gmail dot com.