_ "unsafe" // for go:linkname
)
-// go119MemoryLimitSupport is a feature flag for a number of changes
-// related to the memory limit feature (#48409). Disabling this flag
-// disables those features, as well as the memory limit mechanism,
-// which becomes a no-op.
-const go119MemoryLimitSupport = true
-
const (
// gcGoalUtilization is the goal CPU utilization for
// marking as a fraction of GOMAXPROCS.
// that can accumulate on a P before updating gcController.stackSize.
maxStackScanSlack = 8 << 10
- // memoryLimitHeapGoalHeadroom is the amount of headroom the pacer gives to
- // the heap goal when operating in the memory-limited regime. That is,
- // it'll reduce the heap goal by this many extra bytes off of the base
- // calculation.
- memoryLimitHeapGoalHeadroom = 1 << 20
+ // memoryLimitMinHeapGoalHeadroom is the minimum amount of headroom the
+ // pacer gives to the heap goal when operating in the memory-limited regime.
+ // That is, it'll reduce the heap goal by this many extra bytes off of the
+ // base calculation, at minimum.
+ memoryLimitMinHeapGoalHeadroom = 1 << 20
+
+ // memoryLimitHeapGoalHeadroomPercent is how headroom the memory-limit-based
+ // heap goal should have as a percent of the maximum possible heap goal allowed
+ // to maintain the memory limit.
+ memoryLimitHeapGoalHeadroomPercent = 3
)
// gcController implements the GC pacing controller that determines
// Updated at the end of each GC cycle, in endCycle.
consMark float64
- // consMarkController holds the state for the mark-cons ratio
- // estimation over time.
- //
- // Its purpose is to smooth out noisiness in the computation of
- // consMark; see consMark for details.
- consMarkController piController
+ // lastConsMark is the computed cons/mark value for the previous 4 GC
+ // cycles. Note that this is *not* the last value of consMark, but the
+ // measured cons/mark value in endCycle.
+ lastConsMark [4]float64
// gcPercentHeapGoal is the goal heapLive for when next GC ends derived
// from gcPercent.
func (c *gcControllerState) init(gcPercent int32, memoryLimit int64) {
c.heapMinimum = defaultHeapMinimum
c.triggered = ^uint64(0)
-
- c.consMarkController = piController{
- // Tuned first via the Ziegler-Nichols process in simulation,
- // then the integral time was manually tuned against real-world
- // applications to deal with noisiness in the measured cons/mark
- // ratio.
- kp: 0.9,
- ti: 4.0,
-
- // Set a high reset time in GC cycles.
- // This is inversely proportional to the rate at which we
- // accumulate error from clipping. By making this very high
- // we make the accumulation slow. In general, clipping is
- // OK in our situation, hence the choice.
- //
- // Tune this if we get unintended effects from clipping for
- // a long time.
- tt: 1000,
- min: -1000,
- max: 1000,
- }
-
c.setGCPercent(gcPercent)
c.setMemoryLimit(memoryLimit)
c.commit(true) // No sweep phase in the first GC cycle.
c.fractionalMarkTime.Store(0)
c.idleMarkTime.Store(0)
c.markStartTime = markStartTime
-
- // TODO(mknyszek): This is supposed to be the actual trigger point for the heap, but
- // causes regressions in memory use. The cause is that the PI controller used to smooth
- // the cons/mark ratio measurements tends to flail when using the less accurate precomputed
- // trigger for the cons/mark calculation, and this results in the controller being more
- // conservative about steady-states it tries to find in the future.
- //
- // This conservatism is transient, but these transient states tend to matter for short-lived
- // programs, especially because the PI controller is overdamped, partially because it is
- // configured with a relatively large time constant.
- //
- // Ultimately, I think this is just two mistakes piled on one another: the choice of a swingy
- // smoothing function that recalls a fairly long history (due to its overdamped time constant)
- // coupled with an inaccurate cons/mark calculation. It just so happens this works better
- // today, and it makes it harder to change things in the future.
- //
- // This is described in #53738. Fix this for #53892 by changing back to the actual trigger
- // point and simplifying the smoothing function.
- heapTrigger, heapGoal := c.trigger()
- c.triggered = heapTrigger
+ c.triggered = c.heapLive.Load()
// Compute the background mark utilization goal. In general,
// this may not come out exactly. We round the number of
c.revise()
if debug.gcpacertrace > 0 {
+ heapGoal := c.heapGoal()
assistRatio := c.assistWorkPerByte.Load()
print("pacer: assist ratio=", assistRatio,
" (scan ", gcController.heapScan.Load()>>20, " MB in ",
//
// So this calculation is really:
// (heapLive-trigger) / (assistDuration * procs * (1-utilization)) /
- // (scanWork) / (assistDuration * procs * (utilization+idleUtilization)
+ // (scanWork) / (assistDuration * procs * (utilization+idleUtilization))
//
// Note that because we only care about the ratio, assistDuration and procs cancel out.
scanWork := c.heapScanWork.Load() + c.stackScanWork.Load() + c.globalsScanWork.Load()
currentConsMark := (float64(c.heapLive.Load()-c.triggered) * (utilization + idleUtilization)) /
(float64(scanWork) * (1 - utilization))
- // Update cons/mark controller. The time period for this is 1 GC cycle.
- //
- // This use of a PI controller might seem strange. So, here's an explanation:
- //
- // currentConsMark represents the consMark we *should've* had to be perfectly
- // on-target for this cycle. Given that we assume the next GC will be like this
- // one in the steady-state, it stands to reason that we should just pick that
- // as our next consMark. In practice, however, currentConsMark is too noisy:
- // we're going to be wildly off-target in each GC cycle if we do that.
- //
- // What we do instead is make a long-term assumption: there is some steady-state
- // consMark value, but it's obscured by noise. By constantly shooting for this
- // noisy-but-perfect consMark value, the controller will bounce around a bit,
- // but its average behavior, in aggregate, should be less noisy and closer to
- // the true long-term consMark value, provided its tuned to be slightly overdamped.
- var ok bool
+ // Update our cons/mark estimate. This is the maximum of the value we just computed and the last
+ // 4 cons/mark values we measured. The reason we take the maximum here is to bias a noisy
+ // cons/mark measurement toward fewer assists at the expense of additional GC cycles (starting
+ // earlier).
oldConsMark := c.consMark
- c.consMark, ok = c.consMarkController.next(c.consMark, currentConsMark, 1.0)
- if !ok {
- // The error spiraled out of control. This is incredibly unlikely seeing
- // as this controller is essentially just a smoothing function, but it might
- // mean that something went very wrong with how currentConsMark was calculated.
- // Just reset consMark and keep going.
- c.consMark = 0
+ c.consMark = currentConsMark
+ for i := range c.lastConsMark {
+ if c.lastConsMark[i] > c.consMark {
+ c.consMark = c.lastConsMark[i]
+ }
}
+ copy(c.lastConsMark[:], c.lastConsMark[1:])
+ c.lastConsMark[len(c.lastConsMark)-1] = currentConsMark
if debug.gcpacertrace > 0 {
printlock()
print(c.heapScanWork.Load(), "+", c.stackScanWork.Load(), "+", c.globalsScanWork.Load(), " B work (", c.lastHeapScan+c.lastStackScan.Load()+c.globalsScan.Load(), " B exp.) ")
live := c.heapLive.Load()
print("in ", c.triggered, " B -> ", live, " B (∆goal ", int64(live)-int64(c.lastHeapGoal), ", cons/mark ", oldConsMark, ")")
- if !ok {
- print("[controller reset]")
- }
println()
printunlock()
}
// Run the background mark worker.
gp := node.gp.ptr()
+ trace := traceAcquire()
casgstatus(gp, _Gwaiting, _Grunnable)
- if trace.enabled {
- traceGoUnpark(gp, 0)
+ if trace.ok() {
+ trace.GoUnpark(gp, 0)
+ traceRelease(trace)
}
return gp, now
}
c.triggered = ^uint64(0) // Reset triggered.
// heapLive was updated, so emit a trace event.
- if trace.enabled {
- traceHeapAlloc(bytesMarked)
+ trace := traceAcquire()
+ if trace.ok() {
+ trace.HeapAlloc(bytesMarked)
+ traceRelease(trace)
}
}
func (c *gcControllerState) update(dHeapLive, dHeapScan int64) {
if dHeapLive != 0 {
+ trace := traceAcquire()
live := gcController.heapLive.Add(dHeapLive)
- if trace.enabled {
+ if trace.ok() {
// gcController.heapLive changed.
- traceHeapAlloc(live)
+ trace.HeapAlloc(live)
+ traceRelease(trace)
}
}
if gcBlackenEnabled == 0 {
goal = c.gcPercentHeapGoal.Load()
// Check if the memory-limit-based goal is smaller, and if so, pick that.
- if newGoal := c.memoryLimitHeapGoal(); go119MemoryLimitSupport && newGoal < goal {
+ if newGoal := c.memoryLimitHeapGoal(); newGoal < goal {
goal = newGoal
} else {
// We're not limited by the memory limit goal, so perform a series of
//
// In practice this computation looks like the following:
//
- // memoryLimit - ((mappedReady - heapFree - heapAlloc) + max(mappedReady - memoryLimit, 0)) - memoryLimitHeapGoalHeadroom
- // ^1 ^2 ^3
+ // goal := memoryLimit - ((mappedReady - heapFree - heapAlloc) + max(mappedReady - memoryLimit, 0))
+ // ^1 ^2
+ // goal -= goal / 100 * memoryLimitHeapGoalHeadroomPercent
+ // ^3
//
// Let's break this down.
//
// terms of heap objects, but it takes more than X bytes (e.g. due to fragmentation) to store
// X bytes worth of objects.
//
- // The third term (marker 3) subtracts an additional memoryLimitHeapGoalHeadroom bytes from the
- // heap goal. As the name implies, this is to provide additional headroom in the face of pacing
- // inaccuracies. This is a fixed number of bytes because these inaccuracies disproportionately
- // affect small heaps: as heaps get smaller, the pacer's inputs get fuzzier. Shorter GC cycles
- // and less GC work means noisy external factors like the OS scheduler have a greater impact.
+ // The final adjustment (marker 3) reduces the maximum possible memory limit heap goal by
+ // memoryLimitHeapGoalPercent. As the name implies, this is to provide additional headroom in
+ // the face of pacing inaccuracies, and also to leave a buffer of unscavenged memory so the
+ // allocator isn't constantly scavenging. The reduction amount also has a fixed minimum
+ // (memoryLimitMinHeapGoalHeadroom, not pictured) because the aforementioned pacing inaccuracies
+ // disproportionately affect small heaps: as heaps get smaller, the pacer's inputs get fuzzier.
+ // Shorter GC cycles and less GC work means noisy external factors like the OS scheduler have a
+ // greater impact.
memoryLimit := uint64(c.memoryLimit.Load())
// Compute the goal.
goal := memoryLimit - (nonHeapMemory + overage)
- // Apply some headroom to the goal to account for pacing inaccuracies.
- // Be careful about small limits.
- if goal < memoryLimitHeapGoalHeadroom || goal-memoryLimitHeapGoalHeadroom < memoryLimitHeapGoalHeadroom {
- goal = memoryLimitHeapGoalHeadroom
+ // Apply some headroom to the goal to account for pacing inaccuracies and to reduce
+ // the impact of scavenging at allocation time in response to a high allocation rate
+ // when GOGC=off. See issue #57069. Also, be careful about small limits.
+ headroom := goal / 100 * memoryLimitHeapGoalHeadroomPercent
+ if headroom < memoryLimitMinHeapGoalHeadroom {
+ // Set a fixed minimum to deal with the particularly large effect pacing inaccuracies
+ // have for smaller heaps.
+ headroom = memoryLimitMinHeapGoalHeadroom
+ }
+ if goal < headroom || goal-headroom < headroom {
+ goal = headroom
} else {
- goal = goal - memoryLimitHeapGoalHeadroom
+ goal = goal - headroom
}
// Don't let us go below the live heap. A heap goal below the live heap doesn't make sense.
if goal < c.heapMarked {
// increase in RSS. By capping us at a point >0, we're essentially
// saying that we're OK using more CPU during the GC to prevent
// this growth in RSS.
- triggerLowerBound := uint64(((goal-c.heapMarked)/triggerRatioDen)*minTriggerRatioNum) + c.heapMarked
+ triggerLowerBound := ((goal-c.heapMarked)/triggerRatioDen)*minTriggerRatioNum + c.heapMarked
if minTrigger < triggerLowerBound {
minTrigger = triggerLowerBound
}
// to reflect the costs of a GC with no work to do. With a large heap but
// very little scan work to perform, this gives us exactly as much runway
// as we would need, in the worst case.
- maxTrigger := uint64(((goal-c.heapMarked)/triggerRatioDen)*maxTriggerRatioNum) + c.heapMarked
+ maxTrigger := ((goal-c.heapMarked)/triggerRatioDen)*maxTriggerRatioNum + c.heapMarked
if goal > defaultHeapMinimum && goal-defaultHeapMinimum > maxTrigger {
maxTrigger = goal - defaultHeapMinimum
}
- if maxTrigger < minTrigger {
- maxTrigger = minTrigger
- }
+ maxTrigger = max(maxTrigger, minTrigger)
// Compute the trigger from our bounds and the runway stored by commit.
var trigger uint64
} else {
trigger = goal - runway
}
- if trigger < minTrigger {
- trigger = minTrigger
- }
- if trigger > maxTrigger {
- trigger = maxTrigger
- }
+ trigger = max(trigger, minTrigger)
+ trigger = min(trigger, maxTrigger)
if trigger > goal {
print("trigger=", trigger, " heapGoal=", goal, "\n")
print("minTrigger=", minTrigger, " maxTrigger=", maxTrigger, "\n")
// If we just disabled GC, wait for any concurrent GC mark to
// finish so we always return with no GC running.
if in < 0 {
- gcWaitOnMark(atomic.Load(&work.cycles))
+ gcWaitOnMark(work.cycles.Load())
}
return out
return n
}
-type piController struct {
- kp float64 // Proportional constant.
- ti float64 // Integral time constant.
- tt float64 // Reset time.
-
- min, max float64 // Output boundaries.
-
- // PI controller state.
-
- errIntegral float64 // Integral of the error from t=0 to now.
-
- // Error flags.
- errOverflow bool // Set if errIntegral ever overflowed.
- inputOverflow bool // Set if an operation with the input overflowed.
-}
-
-// next provides a new sample to the controller.
-//
-// input is the sample, setpoint is the desired point, and period is how much
-// time (in whatever unit makes the most sense) has passed since the last sample.
-//
-// Returns a new value for the variable it's controlling, and whether the operation
-// completed successfully. One reason this might fail is if error has been growing
-// in an unbounded manner, to the point of overflow.
-//
-// In the specific case of an error overflow occurs, the errOverflow field will be
-// set and the rest of the controller's internal state will be fully reset.
-func (c *piController) next(input, setpoint, period float64) (float64, bool) {
- // Compute the raw output value.
- prop := c.kp * (setpoint - input)
- rawOutput := prop + c.errIntegral
-
- // Clamp rawOutput into output.
- output := rawOutput
- if isInf(output) || isNaN(output) {
- // The input had a large enough magnitude that either it was already
- // overflowed, or some operation with it overflowed.
- // Set a flag and reset. That's the safest thing to do.
- c.reset()
- c.inputOverflow = true
- return c.min, false
- }
- if output < c.min {
- output = c.min
- } else if output > c.max {
- output = c.max
- }
-
- // Update the controller's state.
- if c.ti != 0 && c.tt != 0 {
- c.errIntegral += (c.kp*period/c.ti)*(setpoint-input) + (period/c.tt)*(output-rawOutput)
- if isInf(c.errIntegral) || isNaN(c.errIntegral) {
- // So much error has accumulated that we managed to overflow.
- // The assumptions around the controller have likely broken down.
- // Set a flag and reset. That's the safest thing to do.
- c.reset()
- c.errOverflow = true
- return c.min, false
- }
- }
- return output, true
-}
-
-// reset resets the controller state, except for controller error flags.
-func (c *piController) reset() {
- c.errIntegral = 0
-}
-
// addIdleMarkWorker attempts to add a new idle mark worker.
//
// If this returns true, the caller must become an idle mark worker unless
// TODO(mknyszek): This isn't really accurate any longer because the heap
// goal is computed dynamically. Still useful to snapshot, but not as useful.
- if trace.enabled {
- traceHeapGoal()
+ trace := traceAcquire()
+ if trace.ok() {
+ trace.HeapGoal()
+ traceRelease(trace)
}
trigger, heapGoal := gcController.trigger()