The task interface to the cluster presents the engines as a fault tolerant, dynamic load-balanced system of workers. Unlike the multiengine interface, in the task interface the user have no direct access to individual engines. By allowing the IPython scheduler to assign work, this interface is simultaneously simpler and more powerful.
Best of all, the user can use both of these interfaces running at the same time to take advantage of their respective strengths. When the user can break up the user’s work into segments that do not depend on previous execution, the task interface is ideal. But it also has more power and flexibility, allowing the user to guide the distribution of jobs, without having to assign tasks to engines explicitly.
To follow along with this tutorial, you will need to start the IPython controller and four IPython engines. The simplest way of doing this is to use the ipcluster command:
$ ipcluster start -n 4
For more detailed information about starting the controller and engines, see our introduction to using IPython for parallel computing.
The first step is to import the IPython IPython.parallel module and then create a Client instance, and we will also be using a LoadBalancedView, here called lview:
In [1]: from IPython.parallel import Client
In [2]: rc = Client()
This form assumes that the controller was started on localhost with default configuration. If not, the location of the controller must be given as an argument to the constructor:
# for a visible LAN controller listening on an external port:
In [2]: rc = Client('tcp://192.168.1.16:10101')
# or to connect with a specific profile you have set up:
In [3]: rc = Client(profile='mpi')
For load-balanced execution, we will make use of a LoadBalancedView object, which can be constructed via the client’s load_balanced_view() method:
In [4]: lview = rc.load_balanced_view() # default load-balanced view
See also
For more information, see the in-depth explanation of Views.
In many cases, you simply want to apply a Python function to a sequence of objects, but in parallel. Like the multiengine interface, these can be implemented via the task interface. The exact same tools can perform these actions in load-balanced ways as well as multiplexed ways: a parallel version of map() and @parallel() function decorator. If one specifies the argument balanced=True, then they are dynamically load balanced. Thus, if the execution time per item varies significantly, you should use the versions in the task interface.
To load-balance map(),simply use a LoadBalancedView:
In [62]: lview.block = True
In [63]: serial_result = map(lambda x:x**10, range(32))
In [64]: parallel_result = lview.map(lambda x:x**10, range(32))
In [65]: serial_result==parallel_result
Out[65]: True
Parallel functions are just like normal function, but they can be called on sequences and in parallel. The multiengine interface provides a decorator that turns any Python function into a parallel function:
In [10]: @lview.parallel()
....: def f(x):
....: return 10.0*x**4
....:
In [11]: f.map(range(32)) # this is done in parallel
Out[11]: [0.0,10.0,160.0,...]
When an AsyncResult object actually maps multiple results (e.g. the AsyncMapResult object), you can actually iterate through them, and act on the results as they arrive:
but want to perform logic on elements in the result, or even abort subsequent
tasks in cases where you are searching for the first affirmative result.
By default, the results will match the ordering of the submitted sequence, but
if you call `map(...ordered=False)`, then results will be provided to the iterator
on a first come first serve basis.
Authors
-------
* MinRK
"""
import time
from IPython import parallel
# create client & view
rc = parallel.Client()
dv = rc[:]
v = rc.load_balanced_view()
# scatter 'id', so id=0,1,2 on engines 0,1,2
dv.scatter('id', rc.ids, flatten=True)
print "Engine IDs: ", dv['id']
# create a Reference to `id`. This will be a different value on each engine
ref = parallel.Reference('id')
See also
When AsyncResult or the AsyncMapResult don’t provide what you need (for instance, handling individual results as they arrive, but with metadata), you can always just split the original result’s msg_ids attribute, and handle them as you like.
For an example of this, see docs/examples/parallel/customresult.py
Often, pure atomic load-balancing is too primitive for your work. In these cases, you may want to associate some kind of Dependency that describes when, where, or whether a task can be run. In IPython, we provide two types of dependencies: Functional Dependencies and Graph Dependencies
Note
It is important to note that the pure ZeroMQ scheduler does not support dependencies, and you will see errors or warnings if you try to use dependencies with the pure scheduler.
Functional dependencies are used to determine whether a given engine is capable of running a particular task. This is implemented via a special Exception class, UnmetDependency, found in IPython.parallel.error. Its use is very simple: if a task fails with an UnmetDependency exception, then the scheduler, instead of relaying the error up to the client like any other error, catches the error, and submits the task to a different engine. This will repeat indefinitely, and a task will never be submitted to a given engine a second time.
You can manually raise the UnmetDependency yourself, but IPython has provided some decorators for facilitating this behavior.
There are two decorators and a class used for functional dependencies:
In [9]: from IPython.parallel import depend, require, dependent
The simplest sort of dependency is requiring that a Python module is available. The @require decorator lets you define a function that will only run on engines where names you specify are importable:
In [10]: @require('numpy', 'zmq')
....: def myfunc():
....: return dostuff()
Now, any time you apply myfunc(), the task will only run on a machine that has numpy and pyzmq available, and when myfunc() is called, numpy and zmq will be imported.
The @depend decorator lets you decorate any function with any other function to evaluate the dependency. The dependency function will be called at the start of the task, and if it returns False, then the dependency will be considered unmet, and the task will be assigned to another engine. If the dependency returns anything other than ``False``, the rest of the task will continue.
In [10]: def platform_specific(plat):
....: import sys
....: return sys.platform == plat
In [11]: @depend(platform_specific, 'darwin')
....: def mactask():
....: do_mac_stuff()
In [12]: @depend(platform_specific, 'nt')
....: def wintask():
....: do_windows_stuff()
In this case, any time you apply mytask, it will only run on an OSX machine. @depend is just like apply, in that it has a @depend(f,*args,**kwargs) signature.
You don’t have to use the decorators on your tasks, if for instance you may want to run tasks with a single function but varying dependencies, you can directly construct the dependent object that the decorators use:
Sometimes you want to restrict the time and/or location to run a given task as a function of the time and/or location of other tasks. This is implemented via a subclass of set, called a Dependency. A Dependency is just a set of msg_ids corresponding to tasks, and a few attributes to guide how to decide when the Dependency has been met.
The switches we provide for interpreting whether a given dependency set has been met:
Sometimes you want to run a task after another, but only if that task succeeded. In this case, success should be True and failure should be False. However sometimes you may not care whether the task succeeds, and always want the second task to run, in which case you should use success=failure=True. The default behavior is to only use successes.
There are other switches for interpretation that are made at the task level. These are specified via keyword arguments to the client’s apply() method.
Note
Dependencies only work within the task scheduler. You cannot instruct a load-balanced task to run after a job submitted via the MUX interface.
The simplest form of Dependencies is with all=True,success=True,failure=False. In these cases, you can skip using Dependency objects, and just pass msg_ids or AsyncResult objects as the follow and after keywords to client.apply():
In [14]: client.block=False
In [15]: ar = lview.apply(f, args, kwargs)
In [16]: ar2 = lview.apply(f2)
In [17]: with lview.temp_flags(after=[ar,ar2]):
....: ar3 = lview.apply(f3)
In [18]: with lview.temp_flags(follow=[ar], timeout=2.5)
....: ar4 = lview.apply(f3)
See also
Some parallel workloads can be described as a Directed Acyclic Graph, or DAG. See DAG Dependencies for an example demonstrating how to use map a NetworkX DAG onto task dependencies.
The schedulers do perform some analysis on graph dependencies to determine whether they are not possible to be met. If the scheduler does discover that a dependency cannot be met, then the task will fail with an ImpossibleDependency error. This way, if the scheduler realized that a task can never be run, it won’t sit indefinitely in the scheduler clogging the pipeline.
The basic cases that are checked:
Warning
This analysis has not been proven to be rigorous, so it is likely possible for tasks to become impossible to run in obscure situations, so a timeout may be a good choice.
Another flag for tasks is retries. This is an integer, specifying how many times a task should be resubmitted after failure. This is useful for tasks that should still run if their engine was shutdown, or may have some statistical chance of failing. The default is to not retry tasks.
Sometimes you may want to re-run a task. This could be because it failed for some reason, and you have fixed the error, or because you want to restore the cluster to an interrupted state. For this, the Client has a rc.resubmit() method. This simply takes one or more msg_ids, and returns an AsyncHubResult for the result(s). You cannot resubmit a task that is pending - only those that have finished, either successful or unsuccessful.
There are a variety of valid ways to determine where jobs should be assigned in a load-balancing situation. In IPython, we support several standard schemes, and even make it easy to define your own. The scheme can be selected via the scheme argument to ipcontroller, or in the TaskScheduler.schemename attribute of a controller config object.
The built-in routing schemes:
To select one of these schemes, simply do:
$ ipcontroller --scheme=<schemename>
for instance:
$ ipcontroller --scheme=lru
lru: Least Recently Used
Always assign work to the least-recently-used engine. A close relative of round-robin, it will be fair with respect to the number of tasks, agnostic with respect to runtime of each task.
plainrandom: Plain Random
Randomly picks an engine on which to run.
twobin: Two-Bin Random
Requires numpy
Pick two engines at random, and use the LRU of the two. This is known to be better than plain random in many cases, but requires a small amount of computation.
leastload: Least Load
This is the default scheme
Always assign tasks to the engine with the fewest outstanding tasks (LRU breaks tie).
weighted: Weighted Two-Bin Random
Requires numpy
Pick two engines at random using the number of outstanding tasks as inverse weights, and use the one with the lower load.
Tasks are assigned greedily as they are submitted. If their dependencies are met, they will be assigned to an engine right away, and multiple tasks can be assigned to an engine at a given time. This limit is set with the TaskScheduler.hwm (high water mark) configurable:
# the most common choices are:
c.TaskSheduler.hwm = 0 # (minimal latency, default)
# or
c.TaskScheduler.hwm = 1 # (most-informed balancing)
The default is 0, or no-limit. That is, there is no limit to the number of tasks that can be outstanding on a given engine. This greatly benefits the latency of execution, because network traffic can be hidden behind computation. However, this means that workload is assigned without knowledge of how long each task might take, and can result in poor load-balancing, particularly for submitting a collection of heterogeneous tasks all at once. You can limit this effect by setting hwm to a positive integer, 1 being maximum load-balancing (a task will never be waiting if there is an idle engine), and any larger number being a compromise between load-balance and latency-hiding.
For maximum throughput, the ‘pure’ scheme is not Python at all, but a C-level MonitoredQueue from PyZMQ, which uses a ZeroMQ DEALER socket to perform all load-balancing. This scheduler does not support any of the advanced features of the Python Scheduler.
Disabled features when using the ZMQ Scheduler:
Task farming will be disabled if an engine unregisters. Further, if an engine is unregistered during computation, the scheduler may not recover.
Since there is no Python logic inside the Scheduler, routing decisions cannot be made based on message content.
The Python schedulers know which engine gets which task, and notify the Hub. This allows graceful handling of Engines coming and going. There is no way to know where ZeroMQ messages have gone, so there is no way to know what tasks are on which engine until they finish. This makes recovery from engine shutdown very difficult.
Note
TODO: performance comparisons
The LoadBalancedView has many more powerful features that allow quite a bit of flexibility in how tasks are defined and run. The next places to look are in the following classes:
The following is an overview of how to use these classes together:
See also
A demo of DAG Dependencies with NetworkX and IPython.