Network Transport in BlazingMQ
This article provides brief overview of the network transport layer of BlazingMQ.
Transport Type
BlazingMQ exclusively uses TCP for all communication within the cluster as well as with clients. TCP provides nice properties of ordering and retransmissions of packets, which makes the implemetation of networking logic simpler in BlazingMQ.
Transport Library
Instead of implementing various low level network building blocks from scratch, BlazingMQ leverages Bloomberg’s Network Transport Framework, which was published as open source in March 2023. NTF enables BlazingMQ to implement non-blocking asynchronous network I/O in an efficient manner. In the words of its authors:
“The Network Transport Framework (NTF) is an open-source collection of libraries for asynchronous network programming for scalable, high-performance applications.
In general, NTF provides building blocks and higher-level abstractions for asynchronously sending and receiving data between processes. More specifically, NTF provides an abstraction around the differences in the native networking APIs offered by the supported operating system, and introduces many useful features commonly required by networked applications. The principle feature of the sockets of this library is the introduction of virtual send and receive queues, which provide a message-oriented, thread-safe (concurrent), send and receive API for sockets using transports with either datagram semantics or stream semantics. Additionally, these virtual queues allow users to operate sockets both reactively (i.e. in the Unix readiness model) or proactively (i.e. in the Windows I/O completion model), regardless of the operating system and interface to that operating system (e.g. select, kqueue, epoll, io_uring, I/O completion ports etc.) NTF supports both stream sockets using TCP and datagram sockets using UDP communicating over either IPv4 or IPv6 networks. NTF also supports local (aka Unix) domain sockets for efficiently communicating between processes on the same machine.
The mechanisms in NTF are scalable, efficient, and multi-threaded. Its libraries are intended for applications needing to manage anywhere from a single socket to many thousands of simultaneous sockets.”
Transport Abstraction
In its implementation, BlazingMQ abstracts the transport by using mwcio interface, which enables BlazingMQ to plug in any implementation which conforms to it. In fact, BlazingMQ was using a legacy network transport library instead of NTF some time back. We have also experimented with plugging in boost::asio as the transport implementation, and using it or another implementation remains a future possibility.
Wire Protocol
BlazingMQ uses a custom wire protocol over TCP, the details of which can be found in this header file. There were four guiding principles when this protocol was designed:
Compactness: The protocol should have minimal overhead.
Forward and backward compatibility: The protocol should be easily extensible while remaining backward compatible.
Batching: The protocol should support batching of messages wherever possible.
Efficiency: The protcol should avoid encoding/decoding overhead for frequent messages, while still taking into consideration concerns like endiannes, etc.
Additionally, several components in the bmqp package provide utilities to build and parse network packets conforming to the wire protocol.