paramsurvey 0.4.19

A toolkit for doing parameter surveys

paramsurvey

paramsurvey is a set of tools for creating and executing parameter surveys, on systems ranging from laptops to supercomputing clusters.

paramsurvey has a pluggable parallel backend. The supported backends at present are python's multiprocessing module, and computing cluster software ray.

Example

import time
import paramsurvey


def sleep_worker(pset, system_kwargs, user_kwargs):
    time.sleep(pset['duration'])
    return {'slept': pset['duration']}


paramsurvey.init(backend='multiprocessing')  # or 'ray', if you installed it

psets = [{'duration': 0.3}] * 5

results = paramsurvey.map(sleep_worker, psets, verbose=2)

for r in results.itertuples():
    print(r.duration, r.slept)
for r in results.iterdicts():
    print(r['duration'], r['slept'])

prints, in addition to some debugging output, a result from each of the 5 sleep_worker calls.

Here are a few more examples:

• The above example, with a few notes
• An example of a multi-stage computation, running several map() functions in a row
• An example of greedy optimization, selecting the best alternative from each map() result
• An example that runs a command-line program for each pset
• An example of using helper functions to create the list of psets to compute
• An example of using the stats to time subsections of your worker function
• An example of using results.missing to re-run missing psets

These examples are installed with the package, so you can run them like this:

$ paramsurvey-readme-example.py
$ paramsurvey-multistage-example.py
$ paramsurvey-greedy-example.py

Philosophy

A parameter survey runs begins by initializing the software, specifying a backend ('multiprocessing' or 'ray').

The user supplies a worker function, which takes a dict of parameters (pset) and returns a dict of results.

The user also supplies a list of parameter sets (psets), perhaps constructed using the helper function paramsurvey.params.product().

Calling pararamsurvey.map() executes the worker function once for each pset. It returns a MapResults object, containing the results, performance statistics, and information about any failures.

You can call paramsurvey.map() more than once.

Keyword arguments to init() and map()

The paramsurvey code has a set of keyword arguments (and corresponding environment variables) to aid debugging and testing. They are:

• backend="multiprocessing" -- which backend to use, currently "multiprocessing" (default) or "ray"
• verbose=1 -- print information about the progress of the computation:
  • 0 = print nothing
  • 1 = print something every 30 seconds (default)
  • 2 = print something every second
  • 3 = print something for every action
• vstats=1 -- controls the verbosity of the performance statistics system, with similar values as verbose
• limit=0 -- limits the number of psets actually computed to this number (0 meaning "all")
• ncores=-1 -- limits the number of cores used, in this case 1 less than the number available (multiprocessing only)
• max_tasks_per_child=3 -- the number of tasks a child will do before restarting. Useful to limit memory leaks. Default: infinite
• group_size=N -- bundle psets into groups, which is useful if a single pset's runtime is too short for ray to efficiently run them. One minute is a good runtime.

Each of these has a corresponding environment variable, e.g. PARAMSURVEY_BACKEND, PARAMSURVEY_VERBOSE. If the environment variable is set, it overrides the values set in the source code. If a kwarg is set for a map() call, that value overrides any value specified for the init() call.

For example, if you wish to debug a large computation by running a small subset of it on a single node, the environment variables allow you to do this without editing your source code:

$ PARAMSURVEY_BACKEND=multiprocessing PARAMSURVEY_VERBOSE=3 PARAMSURVEY_LIMIT=10 ./myprogram.py

For retrospective debugging, i.e. your run crashes and you are sad that you specified a lower verbosity than you desire post-crash, paramsurvey creates a hidden logfile in the current directory for every run, named .paramusurvey-DATE-TIME.log. For example, this hidden logfile will always contain information about any exceptions raised in your worker code.

Pandas-related quirks

To preserve memory in the case of large numbers of psets, the psets and the results are stored as a Pandas DataFrame. This creates a couple of quirks visible to user code:

• A key that exists in any pset or result will exist in every pset or result, with a default value of nan
• If you treat nan as a boolean in Python, it evaluates to True
• Because of a (fixable) quirk in pandas-appender, columns are not automatically promoted from integer to float. So if you have a large number of psets with an integer value, and then throw in a float, it will be rounded to an integer.

Backend-specific arguments

Both init() and map() take a backend-specific keyword argument named for the backend, and ignored by other backends. For example, to pass an argument only used by the ray backend,

paramsurvey.init(..., ray={'address': 'auto'})
...
paramsurvey.map(..., ray={'num_gpus': 1})

Individual psets can also have backend-specific arguments, in the case where different units of work need different memory limits or different thread counts. The complete list of ray arguments is: name, memory, num_cpus, and num_gpus. There is also num_cores as an alias for num_cpus.

The MapResults object

The MapResults object has several properties:

• results is a Pandas DataFrame containing the values of the pset and the keys returned by the worker function. Iterating over these results is documented above, as either dicts or tuples.
• missing is a DataFrame of psets that did not generate results, plus extra '_exception' and '_traceback' columns if an exception was raised in the worker function.
• progress is a MapProgress object with properties containing the details of pset execution: total, active, finished, failures, exceptions.
• stats is a PerfStats object containing performance statistics.

Fault tolerance, and re-running failed computations

If you have bugs in your code that raise exceptions, paramsurvey will diligently collect the exceptions and tracebacks and will print them into the output (verbose=2 or more) and also in the hidden logfile .paramusurvey-DATE-TIME.log mentioned above. Any pset causing an exception will be in the results.missing DataFrame.

In addition to exceptions thrown by user code, the paramsurvey module and the laptop or cluster its running on can experience two kinds of errors. The first is failures of some nodes or processes in one of the distributed backends, like ray, ray will quietly re-run the pset as long as the head node and the driver are still alive. So, for example, it's safe to run most nodes in the computation as a "EC2 spot instance" or "preemptable node" in a cluster queue system like Slurm.

The other kind of error is one that can't be caught by paramsurvey. This might include python multiprocessing completely crashing everything because your laptop is out of memory, or ray having indigestion because some nodes are low on memory and are responding slowly, but aren't totally dead.

Debugging checklist

• Check the hidden logfile for details of previous runs
• Use environment variables to shrink your run to a single pset, as mentioned above:

$ PARAMSURVEY_BACKEND=multiprocessing PARAMSURVEY_VERBOSE=3 PARAMSURVEY_LIMIT=1 ./myprogram.py

• Memory debugging: print(paramsurvey.utils.vmem()) in a few places, values are in GBytes
• Slow memory leaks: restart children after every Nth pset by adding max_tasks_per_child=10 to the paramsurvey.init() call
• Look at the performance statistics
• Look at the example scripts linked above, which demonstrate most features mentioned in this README

Carbon monitoring

If you'd like to monitor your paramsurvey run in real-time on one of those fancy modern dashboard thingies, pass details to paramsurvey.init(..., carbon_server="127.0.0.1", carbon_port=2004, carbon_prefix="paramsurvey") The integrated code only knows how to send pickle and (so far) only records progress.

Worker function limitations

The worker function runs in a different address space and possibly on a different server. It shouldn't access any global variables.

For hard-to-explain Python reasons, define the worker function before calling paramsurvey.init(). The worker function should not be nested inside another function. On Windows, the main program file should have a if __name == '__main__' guard similar to the examples at the top of the Python multprocessing documentation.

Installing

$ pip install paramsurvey
$ pip install paramsurvey[ray]