Add Fiber Running / Fiber Switch events to the marker chart

[tenderlove]

Fiber Running and Fiber Switch events give us information about what Fibers in the system are doing. The Fiber Running span in the marker chart allow us to easily select work that a particular Fiber is doing, where the Fiber Switch event allows us to identify why a Fiber changed.

Here is a sample program we'll profile with Vernier:

require "async"
require "async/queue"

def measure
  x = Process.clock_gettime(Process::CLOCK_MONOTONIC)
  yield
  Process.clock_gettime(Process::CLOCK_MONOTONIC) - x
end

def fib(n)
  if n < 2
    n
  else
    fib(n-2) + fib(n-1)
  end
end

# find fib that takes ~50ms
fib_i = 50.times.find { |i| measure { fib(i) } >= 0.05 }
sleep_i = measure { fib(fib_i) }

Async {
  latch = Async::Queue.new

  workers = [
    Async {
      latch.pop # block until ready to measure

      100.times {
        sleep(sleep_i)
        # stalls happen here. This worker wants to be scheduled so it can
        # continue the loop, but will be blocked by another worker executing fib
      }
    },
    Async {
      latch.pop # block until ready to measure

      100.times { fib(fib_i) }
    },
  ]

  2.times { latch << nil }
  workers.each(&:wait)
}

The sample program has one Fiber that is CPU bound. It's calculating Fib in a loop. The other Fiber is simulating IO work by calling sleep.

Here I've highlighted the marker where we have one Fiber using the CPU for a long time:

Selecting that marker lets us verify that in fact it's spending time calculating Fib:

[jhawthorn]

I realized only after we wrote this 😅, I think we need to scope the pairing of fiber switch events and converting them to the "fiber running" spans by thread.

[jhawthorn]

I refactored a whole bunch so that keeping things scoped by thread wasn't so awkward so 🤞 that didn't break too much