Writing task-parallel-safe steps

Polaris runs one step at a time today, and a step that changes the process working directory or mutates a module-level global is correct under that assumption. It stops being correct as soon as two steps run concurrently in one process, and the failure is timing-dependent: it shows up as intermittent wrong answers rather than as a clean error.

Polaris is adding task parallelism so that analysis steps, which are heavily Python and are the workload that most needs to run concurrently, can be scheduled together and across nodes. The rules below are what a step must satisfy to be scheduled that way. They are cheap to follow when a step is first written and expensive to retrofit, so new analysis steps should follow them from the start. Existing steps are being converted only where they are shared building blocks; see Current status.

The reasoning behind the rules, and the scheduler work they support, is in the design documents, under “Task-Parallel-Safe Analysis Steps in Polaris”.

What Polaris already guarantees

Several properties a distributed executor would otherwise have to demand are already true, so a step does not have to do anything special about them:

Steps are already serializable. polaris setup pickles each step to step.pickle in its work directory and polaris serial unpickles it in a different process. A step holding an open file handle or a live connection would already fail today.
Config options already come from a file. Each step’s config file is written into its work directory at setup and re-read immediately before the step runs, so a step run in a fresh process sees the same config.
Declared inputs and outputs are already absolute. Step.process_inputs_and_outputs() resolves the filename arguments to add_input_file() and add_output_file() against the step’s work directory at setup, so self.inputs and self.outputs do not depend on the process working directory.
Every step already has its own work directory and its own logger.

The rules

Address files through the step, not the working directory

Polaris changes the process working directory to the step’s work directory before calling run(), so a bare relative filename works today. The working directory is process-global, so it cannot stay pointed at your step once another step is running beside you.

Build paths with polaris.Step.work_path() instead:

class Viz(Step):
    def run(self):
        with xr.open_dataset(self.work_path('output.nc')) as ds:
            ...
        fig.savefig(self.work_path('comparison.png'))

This applies to paths handed to subprocesses and to third-party libraries as well, since those inherit or resolve against the same working directory. The filename arguments to add_input_file() and add_output_file() are unaffected – they are relative by design and already resolved at setup.

Do not mutate process-global state

A step must not change anything shared by the whole process: the working directory, environment variables, module-level configuration in libraries such as mpas_tools.io, or global plotting state. In particular:

Do not assign to plt.rcParams or call plt.style.use(). Use the polaris.viz.mplstyle_context() context manager, which applies the Polaris style (and an optional dpi) for the duration of a with block and restores the previous settings afterwards.
Do not call plt.switch_backend(). Matplotlib already selects a non-interactive backend where Polaris steps run, and the object-oriented figure API does not consult the backend at all.

Use the object-oriented figure API

pyplot’s “current figure” is process-global. plt.figure(), plt.subplot(), plt.title(), plt.colorbar() and plt.savefig() all act on whichever figure pyplot made most recently, so two steps plotting at the same time draw into each other’s figures. Build the figure explicitly instead:

from matplotlib.figure import Figure

from polaris.viz import mplstyle_context

with mplstyle_context():
    fig = Figure(figsize=(8, 4.5))
    ax = fig.add_subplot(111)
    ax.plot(x, y)
    ax.set_title('...')
    fig.colorbar(handle, ax=ax)
    fig.savefig(self.work_path('my_plot.png'))

A figure built this way is never registered with pyplot, so there is also nothing to plt.close().

Keep temporary files in the step’s work directory

Write scratch files under the step’s own work directory rather than in /tmp or the base work directory, so that two concurrent steps cannot choose the same path. A step must also not write into another step’s work directory, and must not read a file that a concurrently running step is still writing.

Log through the step’s logger

Emit output with self.logger, never with print() or by writing to stdout, stderr or the root logger. Output written to shared streams interleaves once several steps run at once, and the resulting log cannot be attributed to a step. Pass self.logger to any helper function or subprocess that produces output; see Logging.

Declare the resources you need

A step must declare the resources it needs so that the scheduler can decide how many steps may run at once. Memory is the field that distinguishes analysis steps from most existing Polaris steps: at high resolution an analysis step may need a large fraction of a node, and a scheduler that packs by cores alone will oversubscribe memory.

Size internal parallelism from what you were given

A step that starts its own pool of processes or threads must size it from the resources the step was given, not from the machine. A pool sized to os.cpu_count() oversubscribes the node as soon as a second step runs beside it. Use self.cpus_per_task:

n_workers = min(self.cpus_per_task, len(work_items))

Current status

The framework support exists – polaris.Step.work_path(), polaris.viz.mplstyle_context(), and the object-oriented figure API in polaris.viz – and the shared building blocks have been converted: polaris.viz, polaris.mesh.spherical, and polaris.ocean.convergence.analysis.

One task-level step, polaris.tasks.seaice.single_column.standard_physics.viz.Viz, has been converted as a worked example of what a conforming step looks like: it is small, it plots, and it reads and writes files, so it exercises most of the rules above in one place.

Most other task-level analysis and visualization steps have not been converted; they still open files by bare relative name and use the pyplot current figure. That is correct under today’s one-step-at-a-time execution and will be addressed as the analysis capability is built out. Do not copy those steps as a template for new work.