proposal: runtime: add SchedStats API

[deft-code]

MemStats provides a way to monitor allocation and garbage collection.

We need a similar facility to monitor the Scheduler.

Briefly:

• Total goroutines create
• Current number of of goroutines
• Total number of goroutines scheduled
• Current number of goroutines scheduled
• Total thread starts
• Current number of threads.
• Metrics on the delay between a goroutine being ready and running on a proc.

[minux]

You can get all of these statistics by parsing the tracing output from runtime/trace.

[aclements]

@minux, while true, runtime/trace seems like a pretty high overhead way to collect what amounts to a fairly small amount of information. It's certainly low overhead for what it does, but what it does is much more than what's needed here. The metrics @deft-code wants are primarily intended for continuous monitoring (based on offline conversations), so it needs to be cheap.

[aclements]

Here are the notes on the desired metrics I had from our meeting a while ago:

Ring buffer of sampled duration between entering and exiting runnable state

• With some probability, when a goroutine enters runnable, tag it, and when it exits runnable, add the runnable duration to a ring
• Consumers can do what they want with these samples, including just averaging them, or building distributions.

Four global stats

• Current number of goroutines
• Total number of goroutines ever created
• Current number of runnable goroutines
• Total number of runnable-to-running transitions

Maybe current number of running goroutines

[bradfitz]

Assigning to @aclements to decide what we're willing to support long-term.

[adg]

Ping @aclements

[rsc]

Still up to @aclements.

[rsc]

Ping @aclements. Can you look at this during the release candidate quiet?

[aclements]

@deft-code, do the specific stats I suggested in #15490 (comment) address your needs?

[aclements]

Sorry, I'd lost track of the fact that there was a concrete proposal doc for this: https://github.com/deft-code/proposal/blob/master/design/15490-schedstats.md

@deft-code, could you mail a CL to add this to the go-proposal repository and, once submitted, edit your first post to link to it? Thanks.

[deft-code]

I'll get on top of it.

[aclements]

Thanks!

[gopherbot]

CL https://golang.org/cl/38180 mentions this issue.

[rsc]

Do we need to keep this issue open, or should we accept it?

[aclements]

There's definitely still work to do on how and what exactly the API should expose, but I think it's pretty clear we need to provide some visibility into the scheduler.

[rsc]

Teams inside Google are patching in CL 38180 and getting some experience with it. If others would like to do the same, please do. We'll probably wait until Go 1.10 to decide to add the API officially. Putting the proposal on hold until then.

[aclements]

Update: Some teams inside Google tried CL 38180 for monitoring CPU load and performing load shedding. However, they've found that the stats provided in the CL aren't a good indicator of CPU load. In particular, it seemed that since goroutines are so cheap (compared to, say, using runnable threads as a load indicator), the runnable goroutine count often fluctuated dramatically, even when the system was under normal load. It was common to see a huge number of goroutines newly started or woken that would exit or sleep almost immediately once run. If the load shedder happened to sample the SchedStats during one of these spikes, it would think the system was overloaded.

There may be some other stat that is a more robust indicator of load. For example, maybe smoothing would help: the runtime could provide the time-integral of the runnable count, from which an application could compute the average runnable count over any desired time window. Or it could expose something similar for the running goroutines to give a measure of idle time (once a system was overloaded, this wouldn't be able to tell how overloaded it was).

[RaduBerinde]

We are looking at ways of detecting overload in CockroachDB, and statistics like described in #15490 (comment) or in the proposal would be extremely helpful. Is there any hope of getting something like this in Go any time soon? I can help with the work if folks are in agreement on what to build. As a timid first step, I was thinking of adding a /sched/goroutines:runnable metric to runtime/metrics - any thoughts on that?

Also, is there any more information about the experiments to use the number of runnable goroutines as an overload indicator? I understand the fluctuations are an issue, but wouldn't sampling this info frequently and looking at an average over a reasonable timeframe address that?