|new 0.0.3||Nov 25, 2020|
|0.0.2||Nov 17, 2020|
|0.0.1||Nov 17, 2020|
#232 in Network programming
The Data Distribution Service for real-time systems (DDS) is an Object Management Group (OMG) machine-to-machine connectivity framework that aims to enable scalable, real-time, dependable, high-performance and interoperable data exchanges using a publish–subscribe pattern. DDS addresses the needs of applications like air-traffic control, smart grid management, autonomous vehicles, robotics, transportation systems, power generation, medical devices, simulation and testing, aerospace and defense, and other applications that require real-time data exchange [Wiki].
This is a pure Rust implementation of DDS. We have tried to translate the key ideas of the DDS application interface to Rust concepts, but also follow Rust conventions, so the API is not quite the same as with DDS specification.
This is still work-in-progress. The most immediate goal is to have enough functionality to be able to communicate with ROS2 software.
Some existing DDS implementation use code generation to implement DataReader and DataWriter calsses for each payload type.
We do not rely on code generation, but Rust generic programming instead: There is a generic DataReader and DataWriter, parameterized with the payload type D and a serializer adapter type SA. The Serde library is used for payload data serialization/deserialization.
The payload type D is required to implement
serde::Serialize when used with a DataWriter, and
serde::DeserializeOwned when used with a DataReader. Many existing Rust types and libraries already support Serde, so they are good to go as-is.
In DDS, a WITH_KEY topic contains multiple differerent instances, that are distingushed by a key. The key must be somehow embedded into the data samples. In our implementation, if the payload type D is communicated in a WITH_KEY topic, then D is additionally required to implement trait
Keyed requires one method:
get_key(&self) -> Self::K , which is used to extract a key of an associated type
D. They key type
K must implement trait
Key, which is a combination of pre-existing traits
Eq + PartialEq + PartialOrd + Ord + Hash + Clone + Serialize + DeserializeOwned and no additional methods.
A serializer adapter type SA (wrapper for a Serde data format) is provided for OMG Common Data Representation (CDR), as this is the default serialization format used by DDS/RTPS. It is possible to use another serialization format for the objects communicated over DDS by providing a Serde data format implementation.
The DDS 1.4 specification specifies an object model and a set of APIs for those objects that constitude the DDS specification. The design of these APIs in, e.g., naming conventions and memory management semantics, does not quite fit the Rust world. We have tried to create a design where important DDS ideas are preserved and implemented, but in a manner suitable to Rust. These design compromises should be apprent only on the application-facing API of DDS. The network side is still aiming to be fully interoperable with existing DDS implementations.
The DDS specifies a class hierarchy, which is part of the API. That hierarchy is not necessarily followed, because Rust does not use inheritance and derived classes in the same sense as e.g. C++.
We have tried to follow Rust naming conventions.
DDS provides two alternative methods for waiting arriving data, namely WaitSets and Listeners. We have chosen to replace these by using the non-blocking IO API from mio crate. The DDS DataReader objects can be directly used with the mio
Poll interface. It should be possible to implement oter APIs, such as an async API on top of that.
DDS uses "instance handles", which behave like pointers to objects managed by the DDS implementation. This does not seem to mix well with Rust memory handling, so we have chosen to not implement those.
An instance handle can be used to refer to refer to data values (samples) with a specific key. We have written the API to use directly the key insted, as that seems semantically equivalent.
The list of standard method return codes specifed by DDS (section 18.104.22.168) is modified, in particaular:
OKcode is not used to indicate successful operation. Success or failure is indicated using the standard
TIMEOUTcode is not used. Timeouts should be indicated as
- The generic
ERRORcode should not be used, but a more specific value instead.
NO_DATAis not used. The absence of data should be encoded as
The DDS specification specifies multiple functions to read received data samples out of a DataReader:
read: Accesses deserialized data objects from a DataReader, and marks them read. Same samples can be read again, if already read samples are requested.
take: Like read, but removes returned objects from DataReader, so they cannot be accessed again.
take_w_condition: Read/take samples that match specified condition.
take_next_sample: Read/take next non-previously accessed sample.
take_instance: Read/take samples belonging to a single instance (having the same key).
take_next_instance: Combination of
take_next_instance_w_condition: Combination of
We have decided to not implement all 12 of these. Instead, we implement smaller collection of methods:
read: Borrows data from the DataReader.
take: Moves data from the DataReader.
take_instance: Access samples belonging to a single key.
All of the methods above require a ReadCondition to specifiy which samples to access, but it is very easy to specify "any" codition, i.e. access unconditionally.
There are also methods
take_next_sample , but these are essentially simplification wrappers for read/take.
In addition to these, we also provide a Rust Iterator interface for reading data.
The DDS specification specifies manual memory management in the sense that many object types are created with a
create_ method call and destoryed with a matching
delete_ method call. We have opted to rely on Rust memory management wherever possible, including handling of payload data.
The RTPS implementation used here is derived from rtps-rs.