Skip to content

Extremely slow message creation for large arrays in Python #156

Open
@Flova

Description

@Flova

The message serialization is very slow (> hundred seconds for large point clouds), when setting the data property of a PointCloud2 message in Python by passing a byte string (e.g. by NumPy asbytes()).

This does not happen if the PointCloud2 message constructor is used (see #155 for the inconsistent behavior) or if a array.array is passed (does not work with the constructor because of #176).

Bug report

Required Info:

  • Operating System:
    • Ubuntu 20.04
  • Installation type:
    • apt
  • Version or commit hash:
    • rolling
  • Client library (if applicable):
    • rclpy

Steps to reproduce issue

  • Create a large Numpy Array called nparr
  • Create a matching PointCloud2 message called msg with parameters matching the NumPy array format
  • Set the data msg.data = nparr.tobytes()
  • Time this process and compare it to setting the data directly in the constructor

The second one is much slower:

  • 2.7484822273254395 seconds # byte string set via the constructor
  • 412.08923530578613 seconds # byte string set via the property

Expected behavior

Both should be equally fast

Actual behavior

The second one is much slower, because an additional assertion is performed.
This assertion is skipped in the constructor because of the bug in #155. It therefore performs much better.

The constructor directly casts the data and sends it to the setter

self.data = array.array('B', kwargs.get('data', []))

The setter itself checks if the cast is already done and skips the casing, including the following assertion

assert \
                ((isinstance(value, Sequence) or
                  isinstance(value, Set) or
                  isinstance(value, UserList)) and
                 not isinstance(value, str) and
                 not isinstance(value, UserString) and
                 all(isinstance(v, int) for v in value) and
                 all(val >= 0 and val < 256 for val in value)), \
                "The 'data' field must be a set or sequence and each value of type 'int' and each unsigned integer in [0, 255]"

The following part of the assertion is critical because it massively slows it down by iterating >2x over all values.

all(isinstance(v, int) for v in value)  and
all(val >= 0 and val < 256 for val in value))

This happens only in __debug__ is set, but that is the case for many dev machines, and this effectively crashes the code by freezing for many seconds.

Implementation considerations

Drop the check for byte strings. As it does not make any sense in the case of this datatype ether way.

I honestly don't know how to change this since the templating seems quite complex. Some advice would be appreciated.

Issue moved from ros2/common_interfaces#177

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions