NanoMQ is a ultra low latency messaging kernel. It enables messaging between processes in much the same way as POSIX message queues but at sub-microsecond latencies. NanoMQ uses efficient wait-free ring buffers arranged in a complete graph. Each node can send messages to any other node, receiving nodes needs to exclusively own a CPU core or HyperThread. The ultra low latency can thus be achieved by avoiding context switches.
Just run make. Requires recent GCC. Tests require Google Test.
On my Intel(R) Core(TM)2 Duo CPU T7500 @ 2.20GHz I get an average RTT of 250 ns and a throughput of 13M msg/s for a two node setup with 100 byte messages. It will be interesting to see measurements on multi CPU systems and the latency depending on which cache the cores share.
In high frequency trading (HFT) systems you want to separate feed handlers and order management systems (OMS) from strategy code in order to increase fault tolerancy and support live deployment of bug fixes or new strategies. NanoMQ allows you to separate these parts of a trading system into separate processes while keeping communication latencies to a fraction of a microsecond.
Git repository: http://github.com/rigtorp/nanomq
Free use of this software is granted under the terms of the GNU
General Public License (GPL). For details see the file
included with the NanoMQ distribution.