Parallel Python documentation

Module API
Quick start guide, SMP
Quick start guide, clusters
Quick start guide, clusters with auto-discovery
Advanced guide, clusters
Command line arguments, ppserver.py
Security and secret key

pp 1.5.7 module API

class Server

Parallel Python SMP execution server class

Methods defined here:

__init__(self, ncpus='autodetect', ppservers=(), secret=None, loglevel=30, logstream=<open file '<stderr>', mode 'w'>, restart=False, proto=0): Creates Server instance

ncpus - the number of worker processes to start on the local
        computer, if parameter is omitted it will be set to
        the number of processors in the system
ppservers - list of active parallel python execution servers
        to connect with
secret - passphrase for network connections, if omitted a default
        passphrase will be used. It's highly recommended to use a
        custom passphrase for all network connections.
loglevel - logging level
logstream - log stream destination
restart - whether to restart worker process after each task completion
proto - protocol number for pickle module

With ncpus = 1 all tasks are executed consequently
For the best performance either use the default "autodetect" value
or set ncpus to the total number of processors in the system

destroy(self): Kills local ppworkers and closes open files

get_active_nodes(self): Returns active nodes as a dictionary
[keys - nodes, values - ncpus]

get_ncpus(self): Returns the number of local worker processes (ppworkers)

get_stats(self): Returns job execution statistics as a dictionary

print_stats(self): Prints job execution statistics. Useful for benchmarking on
clusters

set_ncpus(self, ncpus='autodetect'): Sets the number of local worker processes (ppworkers)

ncpus - the number of worker processes, if parammeter is omitted
it will be set to the number of processors in the system

submit(self, func, args=(), depfuncs=(), modules=(), callback=None, callbackargs=(), group='default', globals=None): Submits function to the execution queue

func - function to be executed
args - tuple with arguments of the 'func'
depfuncs - tuple with functions which might be called from 'func'
modules - tuple with module names to import
callback - callback function which will be called with argument
list equal to callbackargs+(result,)
as soon as calculation is done
callbackargs - additional arguments for callback function
group - job group, is used when wait(group) is called to wait for
jobs in a given group to finish
globals - dictionary from which all modules, functions and classes
will be imported, for instance: globals=globals()

wait(self, group=None): Waits for all jobs in a given group to finish.
If group is omitted waits for all jobs to finish

default_port = 60000

default_secret = 'epo20pdosl;dksldkmm'

class Template

Template class

Methods defined here:

__init__(self, job_server, func, depfuncs=(), modules=(), callback=None, callbackargs=(), group='default', globals=None): Creates Template instance

jobs_server - pp server for submitting jobs
func - function to be executed
depfuncs - tuple with functions which might be called from 'func'
modules - tuple with module names to import
callback - callback function which will be called with argument
list equal to callbackargs+(result,)
as soon as calculation is done
callbackargs - additional arguments for callback function
group - job group, is used when wait(group) is called to wait for
jobs in a given group to finish
globals - dictionary from which all modules, functions and classes
will be imported, for instance: globals=globals()

submit(self, *args): Submits function with *arg arguments to the execution queue

Data
		copyright = 'Copyright (c) 2005-2009 Vitalii Vanovschi. All rights reserved' version = '1.5.7'

Quick start guide, SMP

1) Import pp module:

import pp

2) Start pp execution server with the number of workers set to the number of processors in the system

job_server = pp.Server()

3) Submit all the tasks for parallel execution:

f1 = job_server.submit(func1, args1, depfuncs1, modules1)

f2 = job_server.submit(func1, args2, depfuncs1, modules1)

f3 = job_server.submit(func2, args3, depfuncs2, modules2)

...etc...

4) Retrieve the results as needed:

r1 = f1()

r2 = f2()

r3 = f3()

...etc...

To find out how to achieve efficient parallelization with pp please take a look at examples

Quick start guide, clusters

On the nodes

1) Start parallel python execution server on all your remote computational nodes:

node-1> ./ppserver.py

node-2> ./ppserver.py

node-3> ./ppserver.py

On the client

2) Import pp module:

import pp

3) Create a list of all the nodes in your cluster (computers where you've run ppserver.py)

ppservers=("node-1", "node-2", "node-3")

4) Start pp execution server with the number of workers set to the number of processors in the system and list of ppservers to connect with :

job_server = pp.Server(ppservers=ppservers)

5) Submit all the tasks for parallel execution:

f1 = job_server.submit(func1, args1, depfuncs1, modules1)

f2 = job_server.submit(func1, args2, depfuncs1, modules1)

f3 = job_server.submit(func2, args3, depfuncs2, modules2)

...etc...

6) Retrieve the results as needed:

r1 = f1()

r2 = f2()

r3 = f3()

...etc...

To find out how to achieve efficient parallelization with pp please take a look at examples

Quick start guide, clusters with autodiscovery

On the nodes

1) Start parallel python execution server on all your remote computational nodes:

node-1> ./ppserver.py -a

node-2> ./ppserver.py -a

node-3> ./ppserver.py -a

On the client

2) Import pp module:

import pp

3) Set ppservers list to auto-discovery:

ppservers=("*",)

4) Start pp execution server with the number of workers set to the number of processors in the system and list of ppservers to connect with :

job_server = pp.Server(ppservers=ppservers)

5) Submit all the tasks for parallel execution:

f1 = job_server.submit(func1, args1, depfuncs1, modules1)

f2 = job_server.submit(func1, args2, depfuncs1, modules1)

f3 = job_server.submit(func2, args3, depfuncs2, modules2)

...etc...

6) Retrieve the results as needed:

r1 = f1()

r2 = f2()

r3 = f3()

...etc...

To find out how to achieve efficient parallelization with pp please take a look at examples

Advanced guide, clusters

On the nodes

1) Start parallel python execution server on all your remote computational nodes (listen to a given port 35000,
and local network interface only, accept only connections which know correct secret):

node-1> ./ppserver.py -p 35000 -i 192.168.0.101 -s "mysecret"

node-2> ./ppserver.py -p 35000 -i 192.168.0.102 -s "mysecret"

node-3> ./ppserver.py -p 35000 -i 192.168.0.103 -s "mysecret"

On the client

2) Import pp module:

import pp

3) Create a list of all the nodes in your cluster (computers where you've run ppserver.py)

ppservers=("node-1:35000", "node-2:35000", "node-3:35000")

4) Start pp execution server with the number of workers set to the number of processors in the system,
list of ppservers to connect with and secret key to authorize the connection:

job_server = pp.Server(ppservers=ppservers, secret="mysecret")

5) Submit all the tasks for parallel execution:

f1 = job_server.submit(func1, args1, depfuncs1, modules1)

f2 = job_server.submit(func1, args2, depfuncs1, modules1)

f3 = job_server.submit(func2, args3, depfuncs2, modules2)

...etc...

6) Retrieve the results as needed:

r1 = f1()

r2 = f2()

r3 = f3()

...etc...

7) Print the execution statistics:

job_server.print_stats()

To find out how to achieve efficient parallelization with pp please take a look at examples

Command line options, ppserver.py

Usage: ppserver.py [-hdar] [-n proto] [-c config_path] [-i interface] [-b broadcast] [-p port] [-w nworkers] [-s secret] [-t seconds]

Options:
-h                 : this help message
-d                 : debug
-a                 : enable auto-discovery service
-r                 : restart worker process after each task completion
-n proto           : protocol number for pickle module
-c path            : path to config file
-i interface       : interface to listen
-b broadcast       : broadcast address for auto-discovery service
-p port            : port to listen
-w nworkers        : number of workers to start
-s secret          : secret for authentication
-t seconds         : timeout to exit if no connections with clients exist

Security and secret key

Due to the security concerns, in Debian we decided to disable default password authentication, and require to run ppserver with a non-trivial secret key (-s command line argument) which should be paired with the matching secret keyword of PP Server class constructor. An alternative way to set a secret key is by assigning pp_secret variable in the configuration file .pythonrc.py which should be located in the user home directory (please make this file readable and writable only by user). The secret key set in .pythonrc.py could be overridden by command line argument (for ppserver) and secret keyword (for PP Server class constructor). Note that passing the password on the command line allows every user to see it (e.g. using ps(1)), and that running it on an open/untrusted network can be a security problem as traffic, including the password, is not encrypted.