Package go-diameter is an implementation of the Diameter Base Protocol RFC 6733 and a stack for the Go programming language.
The current implementation is solid and works fine for general purpose clients and servers. It can send and receive messages efficiently as well as build and parse AVPs based on dictionaries.
See the API documentation at http://godoc.org/github.com/fiorix/go-diameter
- Comprehensive XML dictionary format
- Embedded dictionaries:
- Base Protocol RFC 6733
- Credit Control RFC 4006
- Network Access Server RFC 7155
- 3GPP specific AVPs from TS 32.299 version 12.7.0
- 3GPP S6a specific commands and AVPs from RFC 5516 and TS 129 272
- Human readable AVP representation (for debugging)
- TLS, IPv4 and IPv6 support for both clients and servers
- Stack based on net/http for simplicity
- Ships with sample client, server, snoop agent and benchmark tool
- State machines for CER/CEA and DWR/DWA for clients and servers
- TCP and SCTP support. SCTP support relies on kernel SCTP implementation and external github.com/ishidawataru/sctp package and is currently tested and enabled for Go 1.8+ and x86 Linux
The easiest way to get started is by trying out the client and server example programs.
With Go 1.11 and newer (preferred), you can start the client and server already:
export GO111MODULE=on
go run github.com/fiorix/go-diameter/v4/examples/server
go run github.com/fiorix/go-diameter/v4/examples/client -hello
Without modules, use standard procedure:
go get github.com/fiorix/go-diameter/examples/...
go run github.com/fiorix/go-diameter/examples/server
go run github.com/fiorix/go-diameter/examples/client -hello
Source code is your best friend. Check out other examples and test cases.
Clients and servers written with the go-diameter package can be quite performant if done well. Besides Go benchmarks, the package ships with a simple benchmark tool to help testing servers and identifying bottlenecks.
In the examples directory, the server has a pprof (http server) that
allows the go pprof
tool to profile the server in real time. The client
can perform benchmarks using the -bench
command line flag.
For better performance, avoid logging diameter messages. Although logging
is very useful for debugging purposes, it kills performance due to a number
of conversions to make messages look pretty. If you run benchmarks on the
example server, make sure to use the -s
(silent) command line switch.
TLS degrades performance a bit, as well as reflection (Unmarshal). Those are important trade offs you might have to consider.
Besides this, the source code (and sub-packages) have function benchmarks that can help you understand what's fast and isn't. You will see that parsing messages is much slower than writing them, for example. This is because in order to parse messages it makes numerous dictionary lookups for AVP types, to be able to decode them. Encoding messages require less lookups and is generally simpler, thus faster.