The goal of this project is to help me understand where things stand with the Erlang VM's scheduler and hard real-time.
Hard real-time means that it's a bug if a computation doesn't complete on time. The Erlang scheduler is a preemptive scheduler with round-robin scheduling within each priority. Hard real-time is NOT a goal. It is possible to modify the Erlang VM to give it more hard real-time properties.
This project starts a process and schedules a timeout event to be sent to it every 1 ms (this is configurable). The process measures the difference between when the event is received and when it should have been received. Zero or more other processes are created that are CPU bound. In a hard real-time system it would be possible to keep the timeout event firing and handling unaffected by the CPU-bound processes.
There are, of course, other considerations when evaluating the scheduler and this project only looks at the performance of one periodic process.
$ cd realtime_tests
$ mix deps.get
$ iex -S mix
iex> RealtimeTests.sweep([cpu_process_flags: [priority: :low]])
The default options sweep through running the test with zero CPU-bound processes and repeats it with more processes. The final printout is a system information dump and CSV data with stats for the min, max, mean and stdev latency. With tweaks to the code, you can dump a histogram of the latencies.
Here's a sample:
System: Erlang/OTP 21 [erts-10.1.1] [source] [64-bit] [smp:4:4] [ds:4:4:10] [async-threads:1] [hipe]
Architecture: x86_64-apple-darwin18.0.0
Input options: []
cpu_workers,min_latency_ms,max_latency_ms,mean_latency_ms,stdev_latency_ms
0,0.016,4.663,1.1939711711711711,1.330280978696712
1,0.023,2.87,1.1572384476895379,1.3052197379418007
2,0.154,10.587,0.8771448289657932,1.0726486196384362
3,0.016,28.502,1.6958215643128625,2.848621327220096
4,0.023,14.236,1.1614870974194837,1.6582175825597263
5,0.03,139.965,4.73500100020004,14.315325723870767
6,0.009,19.773,2.1565601120224045,3.1509438434572146
See LOGS.md for raw data.