0

I am currently working with building realtime IPC system.

I am trying to build a realtime system free of all undeterministic components.

I try to setup IPC based on shared memory model.

Is it safe to use managed_shared_memory from boost library to minimize unpredictable latency? or should I use mmap() for memory sharing?

Also Is rt_signal which has channel from SIGRTMIN to SIGRTMAX safe to use?

Thanks! By the way I am using them on preempt-rt patched linux kernel.

BabySWEngineer
  • 3
  • 1

1 Answers1

0

Is it safe to use managed_shared_memory from boost library to minimize unpredictable latency? or should I use mmap() for memory sharing?

Manipulating process address space mapping with mmap or brk is an expensive operation, so you'd want to minimize the number of times you use mmap. mmap is one source of unpredictable latency.

managed_shared_memory uses mmap and provides extra functionality on top of that, which you'd need to implement yourself when using mmap.

It probably doesn't matter for performance which one you use since mmap cost is going to dominate.

Also Is rt_signal which has channel from SIGRTMIN to SIGRTMAX safe to use?

Signals may not be the best choice for IPC because the POSIX signal interfaces are prone to race conditions: process ids get reused so that pid argument may refer to a totally different process by the time you call sigqueue.

Signals only support directed communication, no broadcast. There is also a system-wide limit on the number of queued real-time signals.

And the APIs to handle signals aren't convenient, but that recently changed with the introduction of signalfd and pidfd_open.

A relatively simple and robust message-based shared-memory IPC is a message ring-buffer in shared memory along with a shared memory mutex and condition variable.

Boost Interprocess library provides interprocess_mutex and interprocess_condition and for ring-buffer you can use boost::lockfree::spsc_queue with a fixed-size.

By the way I am using them on preempt-rt patched linux kernel.

You should watch Kernel Recipes 2016 - Who needs a Real-Time Operating System (Not You!) - Steven Rostedt. Real-time basically means slow but with guaranteed dead-lines.

Maxim Egorushkin
  • 131,725
  • 17
  • 180
  • 271
  • Thanks for your reply. I have a further question. I am looking for a asynchronous event signal that could be used for realtime system. For example, process A updates a piece of shared memory with fresh data, then notifies process B which was waiting for the piece of shared memory that now process B can use it. What would be the fastest (average latency wise) and deadline guaranteeing mechanism for it assuming preempt rt is being used. it would be great if max latency is always less than 30us. – BabySWEngineer Mar 18 '20 at 13:50
  • @BabySWEngineer If you just update a shared structure, then you don't need a queue. The process mutex and condition variable allow you to notify the waiters. It takes roughly 1us to wake up a waiting process using condition variable. – Maxim Egorushkin Mar 18 '20 at 13:55