Performance

• Generally when using Python, you will get the usual Python performance, while when using Node.js, you will get the usual Node.js performance
• Performance is significantly better with Python 3.9 or later because of the fast call protocol
• Calling into Python from JS is more expensive than Python to Python, tests show that using Node.js with Python 3.9+ for numpy arrays:
  • of 8 elements is 4 times slower using raw access and 10 times slower using proxify
  • of 512 elements is 3 times slower using raw access and 4 times slower using proxify
  • of 8192 elements, there is no significant difference with raw access and proxify is 25% slower
  • of 32768 elements, there is no significant difference whatever the access method is
• It is possible to mix raw access and proxify - the underlying PyObject can be accessed as the __PyObject__ from a proxy object
• toJS() and fromJS() are the most expensive parts as they copy objects between the Python and the JavaScript heap
  • For best performance try to keep objects in Python and in JavaScript as much as possible and avoid moving them
• The memory usage of your program will be the sum of the memory usage of a Python interpreter (not that much) and a Node.js interpreter (more significant)
• All Python objects referenced in JavaScript have an additional 56-byte header that is a V8 object - in fact most of the performance loss when processing very small Python objects is related to the creation and the handling of this additional V8 object - that is the V8 tip of the Python iceberg
  • This means that when Python code creates a list, this list stays in the Python heap
  • When this list is seen for the first time by V8, it receives its 56-byte header
  • If the JavaScript code traverses the list accessing every element, then every element receives a 56-byte header - but these are available from immediate garbage collection
  • If every element is separately pushed into a JavaScript array as a PyObject, then all 56-byte headers must be kept
• The two GCs should work very well in tandem