Implementations and benchmarks for various inter-process-communication (IPC) methods on Linux and OS X.
The following IPC methods are implemented. To measure their sequential throughput we send a single message forth and back (i.e., ping pong) between two processes.
| Method | 100 Byte Messages | 1 Kilo Byte Messages |
|---|---|---|
| Unix Signals | --broken-- | --broken-- |
| ZeroMQ (TCP) | 24,901 msg/s | 22,679 msg/s |
| Internet sockets (TCP) | 70,221 msg/s | 67,901 msg/s |
| Domain sockets | 130,372 msg/s | 127,582 msg/s |
| Pipes | 162,441 msg/s | 155,404 msg/s |
| Message Queues | 232,253 msg/s | 213,796 msg/s |
| FIFOs (named pipes) | 265,823 msg/s | 254,880 msg/s |
| Shared Memory | 4,702,557 msg/s | 1,659,291 msg/s |
| Memory-Mapped Files | 5,338,860 msg/s | 1,701,759 msg/s |
NOTE: The code is rather old and there might be sub-optimal configurations!
We are happy to update the configuration with concrete suggestions (see contributions below).
In particular, zeromq is a great library and should probably be performing better, especially because we are only using the TCP implementation.
There is one dedicated for IPC (-> TODO).
In addition, there is little technical reason for shared memory to perform differently than memory-mapped files (could be due to a lack of warmup).
Non-the-less, hopefully, this benchmark can serve as a solid starting point by providing ball-park numbers and a reference implementation.
For a detailed evaluation of inter-node communication, see our L5 library.
Some required packages on Ubuntu:
sudo apt-get install pkg-config, libzmqpp-devYou can build the project and all necessary executables using CMake. The following commands (executed from the root folder) should do the trick:
mkdir build
cd build
cmake ..
makeThis will generate a build/source folder, holding further directories for each IPC type.
Simply execute the program named after the folder, e.g. build/source/shm/shm.
Where applicable, this will start a new server and client process, run benchmarks and print results to stdout. For example, running build/source/shm/shm outputs:
============ RESULTS ================
Message size: 4096
Message count: 1000
Total duration: 1.945 ms
Average duration: 1.418 us
Minimum duration: 0.000 us
Maximum duration: 25.000 us
Standard deviation: 1.282 us
Message rate: 514138 msg/s
=====================================
To control the number of messages sent and the size of each message, each master executable (which stars the server and client) takes two optional command-line arguments:
-c <count>: How many messages to send between the server and client. Defaults to 1000.-s <size>: The size of individual messages. Defaults to 1000.
For example, you can measure the latency of sending 100 bytes a million times via domain sockets with the following command:
$ ./domain -c 1000000 -s 100We also provide a shell script under results/ that runs all methods with various configurations and stores the results.
Some tests may have issues due to system limits, so you may want to re-run the script or run some tests manually.
Contributions are welcome, as long as they fit within the goal of this benchmark: sequential single-node communication.
Just open an issues or send a pull request.
This project is released under the MIT License. For more information, see the LICENSE file.
Peter Goldsborough + cat ❤️
Some maintenance by Alexander van Renen