GopherCon 2018 - Go Scheduler
These are some notes from my experiences at the GopherCon 2018. I don’t expect these will be laid out in any particularly useful way; I am mostly taking them so I can remember some of the bits I found most useful in the future.
Go Scheduler Decisions
- Can support hundreds of thousands of goroutines
- How does the scheduler do what it does in a fair way?
User-space threads managed by the go runtime
- Faster to create, destroy, context switch than OS threads
- Smaller memory footprint
- Need a scheduler to assign to OS threads
- Invoked at goroutine creation and blocking
- Syscalls also, but those also block the underlying OS thread
- Use a small number of OS threads
- Support high concurrency
- Use all possible cores on a box (when appropriate)
How do goroutines get onto OS threads?
- When to create OS threads?
- When/how to distribute?
FIFO runqueue for things that need to be run
Can’t do single-thread, since syscalls will block the OS thread (and no //-ism)
Don’t want to do thread-per-goroutine (expensive)
Create threads when needed, but keep idle ones around to handle future goroutines
Need a lock on the runqueue
Also, what about creating lots of goroutines?
- Lock contention ahoy!
- Limit the number of OS threads – but to what? Number of cores
- BUT! Increasing number of cores gets us back to heavy contention
Maybe split up the runqueue into N runqueues?
- But what happens if the thread’s runqueue is empty?
- Pick another queue at random and steal half the work to do
- What happens if an OS thread blocks on a syscall? Background thread to redistribute work and start another OS thread (since the other is blocked)
What about non-cooperative goroutines (long running high-cpu ops)?
- Can starve the runqueues
- Need mechanism for preemption
- Background thread “sysmon” to detect long-running goroutines and unschedule them, putting them on a lower-priority global runqueue (threads check the global queue less often than their own runqueues)
- No priorities (FIFO)
- No strong preemption
- Not aware of system topology – bad for cache locality