Frequently Asked Questions¶
Does xarray-simlab provide built-in models?¶
No, xarray-simlab provides only the framework for creating, customizing and running computational models. It is intended to be a general-purpose tool. Domain specific models should be implemented in 3rd party packages. For example, xarray-topo provides xarray-simlab models and model components for simulating landscape evolution.
Can xarray-simlab be used with existing model implementations?¶
Yes, it should be easy to wrap existing model implementations using xarray-simlab. Even monolithic codes may leverage the xarray interface. However, as the framework works best at a fine grained level (i.e., with models built from many “small” components) it might be worth to refactor those monolithic implementations.
Does xarray-simlab allow fast model execution?¶
Yes, although it depends on how the model is implemented.
xarray-simlab is written in pure-Python and so is the outer (time) loop in simulations. The execution of Python code is slow compared to other languages, but for the outer loop only it wouldn’t represent the main bottleneck of the overall model execution, especially when using an implicit time scheme. For inner (e.g., spatial) loops in each model processes, it might be better to have a numpy vectorized implementation, use tools like Cython or Numba or call wrapped code that is written in, e.g., C/C++ or Fortran (see for example f2py for wrapping Fortran code or pybind11 for wrapping C++11 code).
As with any other framework, xarray-simlab introduces an overhead compared to a simple, straightforward (but non-flexible) implementation of a model. The preliminary benchmarks that we have run show only a very small (almost free) overhead, though. This overhead is mainly introduced by the thin object-oriented layer that model components (i.e., Python classes) together form.
Does xarray-simlab support running model(s) in parallel?¶
There is currently no support for model execution in parallel but it is a top priority for the next releases!
Three levels of parallelism are possible:
- “inter-model” parallelism, i.e., execution of multiple model runs in parallel,
- “inter-process” parallelism, i.e., execution of multiple processes of a model in parallel,
- “intra-process” parallelism, i.e., parallel execution of some code written in one or more processes.
Note that the notion of process used above is different from multiprocessing: a process here corresponds to a component of a model (see Modeling Framework section).
The first level “inter-model” is an embarrassingly parallel problem. Next versions of xarray-simlab will allow to very easily run simulations in parallel (e.g., for sensitivity analyses).
It shouldn’t be hard to add support for the second level “inter-process” given that processes in a model together form a directed acyclic graph. However, those processes usually perform most of their computation on shared data, which may significantly reduce the gain of parallel execution when using multiple OS processes or in distributed environments. Using multiple threads is limited by the CPython’s GIL, unless it is released by the code executed in model processes.
The third level “intra-process” is more domain specific. Users are free to develop xarray-simlab compatible models with custom code (in processes) that is executed either sequentially or in parallel.
Is it possible to use xarray-simlab without xarray?¶
Although it sounds a bit odd given the name of this package, in principle it is possible. The implementation of the modeling framework is indeed completely decoupled from the xarray interface.
However, the xarray extension provided in this package aims to be the primary, full-featured interface for setting and running simulations from within Python.
The modeling framework itself doesn’t have any built-in interface apart from a few helper functions for running specific stages of a simulation. Any other interface has to be built from scratch, but in many cases it wouldn’t require a lot of effort. In the future, we plan to also provide an experimental interface for real-time, interactive simulation based on tools like ipywidgets, bokeh and/or holoviews.
Will xarray-simlab support Python 2.7.x?¶
No, unless there are very good reasons to do so. The main packages of the Python scientific ecosystem support Python 3.4 or later, and it seems that Python 2.x will not be maintained anymore past 2020 (see PEP 373). Although some tools easily allow supporting both Python 2 and 3 versions in a single code base, it still makes the code harder to maintain.