Doing something as "at the same time" as possible

Questions / Help
,
1

I know beam makes a scheduler per CPU (OS thread); so theoretically I can do something at exactly the same time up to the number of CPUs I have.

What is the best way to kick them all off at the same time? Is my only option to send (via Process.send or Kernel.send) a message to a process running on each scheduler? If I do that, the messages are sent out serially. Any alternative techniques or should that be considered fast enough in this ecosystem?

I posted a similar post, but I think it was too complex to bother with.

2

The type of question you’re asking I think requires some definitions. There is, at the end of the day, no such thing as truly “at the same time” there is “measured to be within a time bound by an observer”. How tight does the time bound need to be? Why? Who is observing this?

3 Likes
3

I think you need to clarify what you mean by “at the same time”, as it’s not well-defined with communicating processes. All it takes is one piece that forces serial behavior to mess things up.

A better approach is to code things to not need guarantees about simultaneity. For instance, in a network-performance situation like your previous post, instead of:

  • start a bunch of connections and a timer simultaneously
  • wait for all the connections to finish
  • read the timer

something like:

  • start a bunch of connections quickly
  • each connection starts a timer when it begins
  • each connection reads its timer at the end
  • each connection sends the timing to the main process
  • the main process waits for all the results, and then averages etc them appropriately
4

You could check how others are doing it.

Maybe you could tell your processes to start at a specific point in time
https://www.erlang.org/doc/man/erlang#monotonic_time-0
then just have them sleep and check time repeatedly.

/one more thing
I don’t think you can choose the specific scheduler a process is on.
:erlang.system_info(:scheduler_id)
I guess if you spawn enough processes under load it should be spread out to all available schedulers.

5

Thanks for the replies, all!

@benwilson512, I think the context you might be seeking is in the linked original post. That post didn’t gent any replies; so I made this post much shorter and simpler. By “at the same time”, I mean parallelism. The goal is to squish that time bound as much as possible in BEAM. I know for other languages the solution would be to start a thread per CPU core.

@al2o3cr, yeah, I think my desires are leaving the realm of practical and entering theoretical (at least in BEAM land). I do have a desire for that TCP SYN to drop in on server/listener to be parallel as possible in BEAM land. Insofar as your recommendation, it’s totally the practical way to do it and may end up what I do if I can’t find what I’m looking for.

@schneebyte, I have seen stressgrid, but not the others. Thanks for the links! I had considered a timer-based kick-off, but wasn’t sure the best way to do it in Elixir/Erlang since I was under the impression that time was pretty squishy in BEAM land.

Regarding choosing specific schedulers, I thought PartitionSupervisor would help partition its children across schedulers. Do I have that messed up in my head?