Tooling

Impact

  • Tooling that gives insight into the state of async runtimes
    • How many tasks are running and what is their state
    • What are tasks blocked on and why?
    • Where is memory allocated?
    • Where is CPU time spent?
    • Flamegraph of where things spend their time
    • Perf-style profile of where things spend their time
  • Tooling should also allow you to limit profiles to a particular request or to requests that meet particular criteria (e.g., coming from a particular source)
  • Tooling should detect common hazards and identify them, suggesting fixes
    • Tasks that clone a Waker but don't trigger it
    • Tasks that don't respond to a request to cancellation for a long time
    • Outlier tasks that sleep for a very long time without being awoken
  • Tooling should permit "always on" profiling that can be used in production
  • Tooling can provide profile-based feedback:
    • Where to "heap-allocate" futures
    • Poll functions that execute for a long time without yielding
    • Imbalanced workloads across cores
  • Tooling can be either customized or integrated into existing tools like perf, gdb, lldb, etc, as appropriate