Billions of Messages Per Minute Over TCP/IP
One of the most important issues when building distributed applications is that of data representation. We must ensure that data sent by a component to a “remote” component (i.e. one that is part of a different process) is received correctly, with the same values. This may seem straightforward but remember that the communicating components may have been written in completely different languages.
Things are complicated further when we consider that different hardware/system architectures are likely to have different ways of representing the “same” values. Simply copying bytes from one component to another is not enough. Even in Java, where we may consider ourselves “protected” from this kind of situation, there is no requirement that two different JVM implementations or different versions from the same vendor use the same internal representation for objects.
The most common solution to this problem is to define a “canonical” representation of data that is understood between processes - even between programming languages - and have data translated into this format before sending and then back into the receiver’s own format once received. Several such “wire formats” exist, ranging from text-based standards such as YAML, JSON or XML, to binary options such as Protobuf that incorporate metadata or are completely raw.
At Chronicle Software we have developed a number of libraries to support the building of applications that are optimised for low latency messaging, primarily in the financial services industry. We provide bespoke solution development and consultancy to clients, the majority of whom are in the financial area, from all over the world.
As software architectures increasingly follow a distributed, event-based approach, we are looking to expand the space in which Chronicle Wire can be used, to support TCP/IP interconnections between components. This article provides a basic overview of the features that will be available and some simple examples of how they can be used.
We are already seeing some startling performance figures for this basic approach - in a benchmark described in Peter Lawrey’s article Java is Very Fast, If You Don't Create Many Objects, for example, which is built upon loopback TCP/IP networking on a single machine, we were able to pass over 4 billion events per minute.
We benchmarked this against similar technologies used for data exchanges, specifically Jackson and BSON. In a test processing 100 byte messages, the 99.99 percentile processing per-message processing time was about 10.5 microseconds with Chronicle Wire compares to 1400 microseconds with Jaskcon/BSON. This is a significant difference.
Here we present an introduction to the key concepts used to realise this. We are, however, designing these features to be flexible as well as performant, and future articles will show some more advanced use cases.
Website link: https://networking-
#protocols #bandwidth #topology #bluetooth
.jpg)
Comments
Post a Comment