skip to navigation
skip to content

Not Logged In

Sheepdog 0.2.0

Shepherd GridEngine

Build Status Coverage Status PyPi Version License

Make Grid Engine a bit more useful from Python.

Requirements

On the host:

On the worker nodes:

  • No requirements beyond the standard library
  • Tested on Python 2.7, Python 3.3
  • Should also work on Python 2.6 and 3.2, 3.4

License

MIT, see LICENSE file.

Overview

Running large map style operations on a Grid Engine cluster can be frustrating. Array jobs can only give scripts an input like some range() function call, but this is rarely sufficient. Collecting results is also a huge pain. Suddenly there are shell scripts and result files everywhere and you feel an overwhelming sense of mediocracy.

Sheepdog aims to make life better for a somewhat specific use case:

  1. You're using Python. Hopefully even Python 3.3.
  2. You've got access to a Grid Engine cluster on some remote machines. They can also run Python, somehow or other. The cluster computers and your client computer can all communicate over a network.
  3. You have a function of several parameters and you want to run it many times with different arguments. Results should come back nicely collated, and are reasonably small (you're not too worried if argument or result objects get copied in memory).
  4. You're a PhD student in Div F at CUED desperately trying to use fear effectively.

To accomplish these aims, Sheepdog:

  1. Takes your function and N tuples of arguments, marshals both
  2. Creates a mapping range(N) to arguments
  3. Starts a network interface (over HTTP)
  4. Starts a size N array job on the Grid Engine cluster, running the client
  5. Each client talks to the server to map its array job ID into an actual set of arguments, and fetches the Python function to execute as well
  6. The function is executed with the arguments
  7. The result is sent back over the network
  8. Results are collated against arguments

This is very similar to:

  • pythongrid. Almost identical. Sheepdog doesn't have to be run on the cluster head, though. And can't resubmit jobs or anything fancy like that. And isn't dead.
  • gridmap. A fork of pythongrid that is actually active and looks quite nice! Maybe look at gridmap.
  • Celery. Yes. Pretty similar.
  • rq. Quite similar.
  • Resque. But Resque is written in Ruby, boo.
  • Every other distributed map compute queue thing ever written.

Usage

Ensure the GridEngine workers have Python available.

Then,

import sheepdog

def f(a, b):
    return a + b

args = [(1, 1), (1, 2), (2, 2)]
config = {"host": "fear"}

results = sheepdog.map_sync(f, args, config)
print("Received results:", results)
# Received results: [2, 3, 4]

There is also support for transferring other functions and variables (using the namespace parameter ns of map_sync) and imports can be handled using global, for example:

def f(a, b):
    import numpy as np
    global np
    return g(a, b)

def g(a, b):
    return np.array((a, b)) ** 2

args = [(1, 2), (3, 4)]
namespace = {"g": g}
config = {"host": "fear"}

results = sheepdog.map_sync(f, args, config, namespace)

See the documentation for full details.

Documentation

View Sheepdog on ReadTheDocs.

 
File Type Py Version Uploaded on Size
Sheepdog-0.2.0.tar.gz (md5) Source 2014-04-18 19KB
  • Downloads (All Versions):
  • 78 downloads in the last day
  • 921 downloads in the last week
  • 3626 downloads in the last month