1 // Copyright 2021 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
9 "internal/goexperiment"
10 "runtime/internal/atomic"
15 // gcGoalUtilization is the goal CPU utilization for
16 // marking as a fraction of GOMAXPROCS.
17 gcGoalUtilization = goexperiment.PacerRedesignInt*gcBackgroundUtilization +
18 (1-goexperiment.PacerRedesignInt)*(gcBackgroundUtilization+0.05)
20 // gcBackgroundUtilization is the fixed CPU utilization for background
21 // marking. It must be <= gcGoalUtilization. The difference between
22 // gcGoalUtilization and gcBackgroundUtilization will be made up by
23 // mark assists. The scheduler will aim to use within 50% of this
26 // Setting this to < gcGoalUtilization avoids saturating the trigger
27 // feedback controller when there are no assists, which allows it to
28 // better control CPU and heap growth. However, the larger the gap,
29 // the more mutator assists are expected to happen, which impact
32 // If goexperiment.PacerRedesign, the trigger feedback controller
33 // is replaced with an estimate of the mark/cons ratio that doesn't
34 // have the same saturation issues, so this is set equal to
36 gcBackgroundUtilization = 0.25
38 // gcCreditSlack is the amount of scan work credit that can
39 // accumulate locally before updating gcController.heapScanWork and,
40 // optionally, gcController.bgScanCredit. Lower values give a more
41 // accurate assist ratio and make it more likely that assists will
42 // successfully steal background credit. Higher values reduce memory
46 // gcAssistTimeSlack is the nanoseconds of mutator assist time that
47 // can accumulate on a P before updating gcController.assistTime.
48 gcAssistTimeSlack = 5000
50 // gcOverAssistWork determines how many extra units of scan work a GC
51 // assist does when an assist happens. This amortizes the cost of an
52 // assist by pre-paying for this many bytes of future allocations.
53 gcOverAssistWork = 64 << 10
55 // defaultHeapMinimum is the value of heapMinimum for GOGC==100.
56 defaultHeapMinimum = (goexperiment.HeapMinimum512KiBInt)*(512<<10) +
57 (1-goexperiment.HeapMinimum512KiBInt)*(4<<20)
59 // scannableStackSizeSlack is the bytes of stack space allocated or freed
60 // that can accumulate on a P before updating gcController.stackSize.
61 scannableStackSizeSlack = 8 << 10
65 if offset := unsafe.Offsetof(gcController.heapLive); offset%8 != 0 {
67 throw("gcController.heapLive not aligned to 8 bytes")
71 // gcController implements the GC pacing controller that determines
72 // when to trigger concurrent garbage collection and how much marking
73 // work to do in mutator assists and background marking.
75 // It uses a feedback control algorithm to adjust the gcController.trigger
76 // trigger based on the heap growth and GC CPU utilization each cycle.
77 // This algorithm optimizes for heap growth to match GOGC and for CPU
78 // utilization between assist and background marking to be 25% of
79 // GOMAXPROCS. The high-level design of this algorithm is documented
80 // at https://golang.org/s/go15gcpacing.
82 // All fields of gcController are used only during a single mark
84 var gcController gcControllerState
86 type gcControllerState struct {
88 // Initialized from GOGC. GOGC=off means no GC.
89 gcPercent atomic.Int32
91 _ uint32 // padding so following 64-bit values are 8-byte aligned
93 // heapMinimum is the minimum heap size at which to trigger GC.
94 // For small heaps, this overrides the usual GOGC*live set rule.
96 // When there is a very small live set but a lot of allocation, simply
97 // collecting when the heap reaches GOGC*live results in many GC
98 // cycles and high total per-GC overhead. This minimum amortizes this
99 // per-GC overhead while keeping the heap reasonably small.
101 // During initialization this is set to 4MB*GOGC/100. In the case of
102 // GOGC==0, this will set heapMinimum to 0, resulting in constant
103 // collection even when the heap size is small, which is useful for
107 // triggerRatio is the heap growth ratio that triggers marking.
109 // E.g., if this is 0.6, then GC should start when the live
110 // heap has reached 1.6 times the heap size marked by the
111 // previous cycle. This should be ≤ GOGC/100 so the trigger
112 // heap size is less than the goal heap size. This is set
113 // during mark termination for the next cycle's trigger.
115 // Protected by mheap_.lock or a STW.
117 // Used if !goexperiment.PacerRedesign.
120 // trigger is the heap size that triggers marking.
122 // When heapLive ≥ trigger, the mark phase will start.
123 // This is also the heap size by which proportional sweeping
126 // This is computed from triggerRatio during mark termination
127 // for the next cycle's trigger.
129 // Protected by mheap_.lock or a STW.
132 // consMark is the estimated per-CPU consMark ratio for the application.
134 // It represents the ratio between the application's allocation
135 // rate, as bytes allocated per CPU-time, and the GC's scan rate,
136 // as bytes scanned per CPU-time.
137 // The units of this ratio are (B / cpu-ns) / (B / cpu-ns).
139 // At a high level, this value is computed as the bytes of memory
140 // allocated (cons) per unit of scan work completed (mark) in a GC
141 // cycle, divided by the CPU time spent on each activity.
143 // Updated at the end of each GC cycle, in endCycle.
145 // For goexperiment.PacerRedesign.
148 // consMarkController holds the state for the mark-cons ratio
149 // estimation over time.
151 // Its purpose is to smooth out noisiness in the computation of
152 // consMark; see consMark for details.
154 // For goexperiment.PacerRedesign.
155 consMarkController piController
157 // heapGoal is the goal heapLive for when next GC ends.
158 // Set to ^uint64(0) if disabled.
160 // Read and written atomically, unless the world is stopped.
163 // lastHeapGoal is the value of heapGoal for the previous GC.
164 // Note that this is distinct from the last value heapGoal had,
165 // because it could change if e.g. gcPercent changes.
167 // Read and written with the world stopped or with mheap_.lock held.
170 // heapLive is the number of bytes considered live by the GC.
171 // That is: retained by the most recent GC plus allocated
172 // since then. heapLive ≤ memstats.heapAlloc, since heapAlloc includes
173 // unmarked objects that have not yet been swept (and hence goes up as we
174 // allocate and down as we sweep) while heapLive excludes these
175 // objects (and hence only goes up between GCs).
177 // This is updated atomically without locking. To reduce
178 // contention, this is updated only when obtaining a span from
179 // an mcentral and at this point it counts all of the
180 // unallocated slots in that span (which will be allocated
181 // before that mcache obtains another span from that
182 // mcentral). Hence, it slightly overestimates the "true" live
183 // heap size. It's better to overestimate than to
184 // underestimate because 1) this triggers the GC earlier than
185 // necessary rather than potentially too late and 2) this
186 // leads to a conservative GC rate rather than a GC rate that
187 // is potentially too low.
189 // Reads should likewise be atomic (or during STW).
191 // Whenever this is updated, call traceHeapAlloc() and
192 // this gcControllerState's revise() method.
195 // heapScan is the number of bytes of "scannable" heap. This
196 // is the live heap (as counted by heapLive), but omitting
197 // no-scan objects and no-scan tails of objects.
199 // For !goexperiment.PacerRedesign: Whenever this is updated,
200 // call this gcControllerState's revise() method. It is read
201 // and written atomically or with the world stopped.
203 // For goexperiment.PacerRedesign: This value is fixed at the
204 // start of a GC cycle, so during a GC cycle it is safe to
205 // read without atomics, and it represents the maximum scannable
209 // lastHeapScan is the number of bytes of heap that were scanned
210 // last GC cycle. It is the same as heapMarked, but only
211 // includes the "scannable" parts of objects.
213 // Updated when the world is stopped.
216 // stackScan is a snapshot of scannableStackSize taken at each GC
217 // STW pause and is used in pacing decisions.
219 // Updated only while the world is stopped.
222 // scannableStackSize is the amount of allocated goroutine stack space in
223 // use by goroutines.
225 // This number tracks allocated goroutine stack space rather than used
226 // goroutine stack space (i.e. what is actually scanned) because used
227 // goroutine stack space is much harder to measure cheaply. By using
228 // allocated space, we make an overestimate; this is OK, it's better
229 // to conservatively overcount than undercount.
231 // Read and updated atomically.
232 scannableStackSize uint64
234 // globalsScan is the total amount of global variable space
235 // that is scannable.
237 // Read and updated atomically.
240 // heapMarked is the number of bytes marked by the previous
241 // GC. After mark termination, heapLive == heapMarked, but
242 // unlike heapLive, heapMarked does not change until the
243 // next mark termination.
246 // heapScanWork is the total heap scan work performed this cycle.
247 // stackScanWork is the total stack scan work performed this cycle.
248 // globalsScanWork is the total globals scan work performed this cycle.
250 // These are updated atomically during the cycle. Updates occur in
251 // bounded batches, since they are both written and read
252 // throughout the cycle. At the end of the cycle, heapScanWork is how
253 // much of the retained heap is scannable.
255 // Currently these are measured in bytes. For most uses, this is an
256 // opaque unit of work, but for estimation the definition is important.
258 // Note that stackScanWork includes all allocated space, not just the
259 // size of the stack itself, mirroring stackSize.
261 // For !goexperiment.PacerRedesign, stackScanWork and globalsScanWork
263 heapScanWork atomic.Int64
264 stackScanWork atomic.Int64
265 globalsScanWork atomic.Int64
267 // bgScanCredit is the scan work credit accumulated by the
268 // concurrent background scan. This credit is accumulated by
269 // the background scan and stolen by mutator assists. This is
270 // updated atomically. Updates occur in bounded batches, since
271 // it is both written and read throughout the cycle.
274 // assistTime is the nanoseconds spent in mutator assists
275 // during this cycle. This is updated atomically. Updates
276 // occur in bounded batches, since it is both written and read
277 // throughout the cycle.
280 // dedicatedMarkTime is the nanoseconds spent in dedicated
281 // mark workers during this cycle. This is updated atomically
282 // at the end of the concurrent mark phase.
283 dedicatedMarkTime int64
285 // fractionalMarkTime is the nanoseconds spent in the
286 // fractional mark worker during this cycle. This is updated
287 // atomically throughout the cycle and will be up-to-date if
288 // the fractional mark worker is not currently running.
289 fractionalMarkTime int64
291 // idleMarkTime is the nanoseconds spent in idle marking
292 // during this cycle. This is updated atomically throughout
296 // markStartTime is the absolute start time in nanoseconds
297 // that assists and background mark workers started.
300 // dedicatedMarkWorkersNeeded is the number of dedicated mark
301 // workers that need to be started. This is computed at the
302 // beginning of each cycle and decremented atomically as
303 // dedicated mark workers get started.
304 dedicatedMarkWorkersNeeded int64
306 // assistWorkPerByte is the ratio of scan work to allocated
307 // bytes that should be performed by mutator assists. This is
308 // computed at the beginning of each cycle and updated every
309 // time heapScan is updated.
310 assistWorkPerByte atomic.Float64
312 // assistBytesPerWork is 1/assistWorkPerByte.
314 // Note that because this is read and written independently
315 // from assistWorkPerByte users may notice a skew between
316 // the two values, and such a state should be safe.
317 assistBytesPerWork atomic.Float64
319 // fractionalUtilizationGoal is the fraction of wall clock
320 // time that should be spent in the fractional mark worker on
321 // each P that isn't running a dedicated worker.
323 // For example, if the utilization goal is 25% and there are
324 // no dedicated workers, this will be 0.25. If the goal is
325 // 25%, there is one dedicated worker, and GOMAXPROCS is 5,
326 // this will be 0.05 to make up the missing 5%.
328 // If this is zero, no fractional workers are needed.
329 fractionalUtilizationGoal float64
331 // test indicates that this is a test-only copy of gcControllerState.
337 func (c *gcControllerState) init(gcPercent int32) {
338 c.heapMinimum = defaultHeapMinimum
340 if goexperiment.PacerRedesign {
341 c.consMarkController = piController{
342 // Tuned first via the Ziegler-Nichols process in simulation,
343 // then the integral time was manually tuned against real-world
344 // applications to deal with noisiness in the measured cons/mark
349 // Set a high reset time in GC cycles.
350 // This is inversely proportional to the rate at which we
351 // accumulate error from clipping. By making this very high
352 // we make the accumulation slow. In general, clipping is
353 // OK in our situation, hence the choice.
355 // Tune this if we get unintended effects from clipping for
362 // Set a reasonable initial GC trigger.
363 c.triggerRatio = 7 / 8.0
365 // Fake a heapMarked value so it looks like a trigger at
366 // heapMinimum is the appropriate growth from heapMarked.
367 // This will go into computing the initial GC goal.
368 c.heapMarked = uint64(float64(c.heapMinimum) / (1 + c.triggerRatio))
371 // This will also compute and set the GC trigger and goal.
372 c.setGCPercent(gcPercent)
375 // startCycle resets the GC controller's state and computes estimates
376 // for a new GC cycle. The caller must hold worldsema and the world
378 func (c *gcControllerState) startCycle(markStartTime int64, procs int) {
379 c.heapScanWork.Store(0)
380 c.stackScanWork.Store(0)
381 c.globalsScanWork.Store(0)
384 c.dedicatedMarkTime = 0
385 c.fractionalMarkTime = 0
387 c.markStartTime = markStartTime
388 c.stackScan = atomic.Load64(&c.scannableStackSize)
390 // Ensure that the heap goal is at least a little larger than
391 // the current live heap size. This may not be the case if GC
392 // start is delayed or if the allocation that pushed gcController.heapLive
393 // over trigger is large or if the trigger is really close to
394 // GOGC. Assist is proportional to this distance, so enforce a
395 // minimum distance, even if it means going over the GOGC goal
397 if goexperiment.PacerRedesign {
398 if c.heapGoal < c.heapLive+64<<10 {
399 c.heapGoal = c.heapLive + 64<<10
402 if c.heapGoal < c.heapLive+1<<20 {
403 c.heapGoal = c.heapLive + 1<<20
407 // Compute the background mark utilization goal. In general,
408 // this may not come out exactly. We round the number of
409 // dedicated workers so that the utilization is closest to
410 // 25%. For small GOMAXPROCS, this would introduce too much
411 // error, so we add fractional workers in that case.
412 totalUtilizationGoal := float64(procs) * gcBackgroundUtilization
413 c.dedicatedMarkWorkersNeeded = int64(totalUtilizationGoal + 0.5)
414 utilError := float64(c.dedicatedMarkWorkersNeeded)/totalUtilizationGoal - 1
415 const maxUtilError = 0.3
416 if utilError < -maxUtilError || utilError > maxUtilError {
417 // Rounding put us more than 30% off our goal. With
418 // gcBackgroundUtilization of 25%, this happens for
419 // GOMAXPROCS<=3 or GOMAXPROCS=6. Enable fractional
420 // workers to compensate.
421 if float64(c.dedicatedMarkWorkersNeeded) > totalUtilizationGoal {
422 // Too many dedicated workers.
423 c.dedicatedMarkWorkersNeeded--
425 c.fractionalUtilizationGoal = (totalUtilizationGoal - float64(c.dedicatedMarkWorkersNeeded)) / float64(procs)
427 c.fractionalUtilizationGoal = 0
430 // In STW mode, we just want dedicated workers.
431 if debug.gcstoptheworld > 0 {
432 c.dedicatedMarkWorkersNeeded = int64(procs)
433 c.fractionalUtilizationGoal = 0
437 for _, p := range allp {
439 p.gcFractionalMarkTime = 0
442 // Compute initial values for controls that are updated
443 // throughout the cycle.
446 if debug.gcpacertrace > 0 {
447 assistRatio := c.assistWorkPerByte.Load()
448 print("pacer: assist ratio=", assistRatio,
449 " (scan ", gcController.heapScan>>20, " MB in ",
450 work.initialHeapLive>>20, "->",
451 c.heapGoal>>20, " MB)",
452 " workers=", c.dedicatedMarkWorkersNeeded,
453 "+", c.fractionalUtilizationGoal, "\n")
457 // revise updates the assist ratio during the GC cycle to account for
458 // improved estimates. This should be called whenever gcController.heapScan,
459 // gcController.heapLive, or gcController.heapGoal is updated. It is safe to
460 // call concurrently, but it may race with other calls to revise.
462 // The result of this race is that the two assist ratio values may not line
463 // up or may be stale. In practice this is OK because the assist ratio
464 // moves slowly throughout a GC cycle, and the assist ratio is a best-effort
465 // heuristic anyway. Furthermore, no part of the heuristic depends on
466 // the two assist ratio values being exact reciprocals of one another, since
467 // the two values are used to convert values from different sources.
469 // The worst case result of this raciness is that we may miss a larger shift
470 // in the ratio (say, if we decide to pace more aggressively against the
471 // hard heap goal) but even this "hard goal" is best-effort (see #40460).
472 // The dedicated GC should ensure we don't exceed the hard goal by too much
473 // in the rare case we do exceed it.
475 // It should only be called when gcBlackenEnabled != 0 (because this
476 // is when assists are enabled and the necessary statistics are
478 func (c *gcControllerState) revise() {
479 gcPercent := c.gcPercent.Load()
481 // If GC is disabled but we're running a forced GC,
482 // act like GOGC is huge for the below calculations.
485 live := atomic.Load64(&c.heapLive)
486 scan := atomic.Load64(&c.heapScan)
487 work := c.heapScanWork.Load() + c.stackScanWork.Load() + c.globalsScanWork.Load()
489 // Assume we're under the soft goal. Pace GC to complete at
490 // heapGoal assuming the heap is in steady-state.
491 heapGoal := int64(atomic.Load64(&c.heapGoal))
493 var scanWorkExpected int64
494 if goexperiment.PacerRedesign {
495 // The expected scan work is computed as the amount of bytes scanned last
496 // GC cycle, plus our estimate of stacks and globals work for this cycle.
497 scanWorkExpected = int64(c.lastHeapScan + c.stackScan + c.globalsScan)
499 // maxScanWork is a worst-case estimate of the amount of scan work that
500 // needs to be performed in this GC cycle. Specifically, it represents
501 // the case where *all* scannable memory turns out to be live.
502 maxScanWork := int64(scan + c.stackScan + c.globalsScan)
503 if work > scanWorkExpected {
504 // We've already done more scan work than expected. Because our expectation
505 // is based on a steady-state scannable heap size, we assume this means our
506 // heap is growing. Compute a new heap goal that takes our existing runway
507 // computed for scanWorkExpected and extrapolates it to maxScanWork, the worst-case
508 // scan work. This keeps our assist ratio stable if the heap continues to grow.
510 // The effect of this mechanism is that assists stay flat in the face of heap
511 // growths. It's OK to use more memory this cycle to scan all the live heap,
512 // because the next GC cycle is inevitably going to use *at least* that much
514 extHeapGoal := int64(float64(heapGoal-int64(c.trigger))/float64(scanWorkExpected)*float64(maxScanWork)) + int64(c.trigger)
515 scanWorkExpected = maxScanWork
517 // hardGoal is a hard limit on the amount that we're willing to push back the
518 // heap goal, and that's twice the heap goal (i.e. if GOGC=100 and the heap and/or
519 // stacks and/or globals grow to twice their size, this limits the current GC cycle's
520 // growth to 4x the original live heap's size).
522 // This maintains the invariant that we use no more memory than the next GC cycle
524 hardGoal := int64((1.0 + float64(gcPercent)/100.0) * float64(heapGoal))
525 if extHeapGoal > hardGoal {
526 extHeapGoal = hardGoal
528 heapGoal = extHeapGoal
530 if int64(live) > heapGoal {
531 // We're already past our heap goal, even the extrapolated one.
532 // Leave ourselves some extra runway, so in the worst case we
533 // finish by that point.
534 const maxOvershoot = 1.1
535 heapGoal = int64(float64(heapGoal) * maxOvershoot)
537 // Compute the upper bound on the scan work remaining.
538 scanWorkExpected = maxScanWork
541 // Compute the expected scan work remaining.
543 // This is estimated based on the expected
544 // steady-state scannable heap. For example, with
545 // GOGC=100, only half of the scannable heap is
546 // expected to be live, so that's what we target.
548 // (This is a float calculation to avoid overflowing on
550 scanWorkExpected = int64(float64(scan) * 100 / float64(100+gcPercent))
551 if int64(live) > heapGoal || work > scanWorkExpected {
552 // We're past the soft goal, or we've already done more scan
553 // work than we expected. Pace GC so that in the worst case it
554 // will complete by the hard goal.
555 const maxOvershoot = 1.1
556 heapGoal = int64(float64(heapGoal) * maxOvershoot)
558 // Compute the upper bound on the scan work remaining.
559 scanWorkExpected = int64(scan)
563 // Compute the remaining scan work estimate.
565 // Note that we currently count allocations during GC as both
566 // scannable heap (heapScan) and scan work completed
567 // (scanWork), so allocation will change this difference
568 // slowly in the soft regime and not at all in the hard
570 scanWorkRemaining := scanWorkExpected - work
571 if scanWorkRemaining < 1000 {
572 // We set a somewhat arbitrary lower bound on
573 // remaining scan work since if we aim a little high,
574 // we can miss by a little.
576 // We *do* need to enforce that this is at least 1,
577 // since marking is racy and double-scanning objects
578 // may legitimately make the remaining scan work
579 // negative, even in the hard goal regime.
580 scanWorkRemaining = 1000
583 // Compute the heap distance remaining.
584 heapRemaining := heapGoal - int64(live)
585 if heapRemaining <= 0 {
586 // This shouldn't happen, but if it does, avoid
587 // dividing by zero or setting the assist negative.
591 // Compute the mutator assist ratio so by the time the mutator
592 // allocates the remaining heap bytes up to heapGoal, it will
593 // have done (or stolen) the remaining amount of scan work.
594 // Note that the assist ratio values are updated atomically
595 // but not together. This means there may be some degree of
596 // skew between the two values. This is generally OK as the
597 // values shift relatively slowly over the course of a GC
599 assistWorkPerByte := float64(scanWorkRemaining) / float64(heapRemaining)
600 assistBytesPerWork := float64(heapRemaining) / float64(scanWorkRemaining)
601 c.assistWorkPerByte.Store(assistWorkPerByte)
602 c.assistBytesPerWork.Store(assistBytesPerWork)
605 // endCycle computes the trigger ratio (!goexperiment.PacerRedesign)
606 // or the consMark estimate (goexperiment.PacerRedesign) for the next cycle.
607 // Returns the trigger ratio if application, or 0 (goexperiment.PacerRedesign).
608 // userForced indicates whether the current GC cycle was forced
609 // by the application.
610 func (c *gcControllerState) endCycle(now int64, procs int, userForced bool) float64 {
611 // Record last heap goal for the scavenger.
612 // We'll be updating the heap goal soon.
613 gcController.lastHeapGoal = gcController.heapGoal
615 // Compute the duration of time for which assists were turned on.
616 assistDuration := now - c.markStartTime
618 // Assume background mark hit its utilization goal.
619 utilization := gcBackgroundUtilization
620 // Add assist utilization; avoid divide by zero.
621 if assistDuration > 0 {
622 utilization += float64(c.assistTime) / float64(assistDuration*int64(procs))
625 if goexperiment.PacerRedesign {
626 if c.heapLive <= c.trigger {
627 // Shouldn't happen, but let's be very safe about this in case the
628 // GC is somehow extremely short.
630 // In this case though, the only reasonable value for c.heapLive-c.trigger
631 // would be 0, which isn't really all that useful, i.e. the GC was so short
632 // that it didn't matter.
634 // Ignore this case and don't update anything.
637 idleUtilization := 0.0
638 if assistDuration > 0 {
639 idleUtilization = float64(c.idleMarkTime) / float64(assistDuration*int64(procs))
641 // Determine the cons/mark ratio.
643 // The units we want for the numerator and denominator are both B / cpu-ns.
644 // We get this by taking the bytes allocated or scanned, and divide by the amount of
645 // CPU time it took for those operations. For allocations, that CPU time is
647 // assistDuration * procs * (1 - utilization)
649 // Where utilization includes just background GC workers and assists. It does *not*
650 // include idle GC work time, because in theory the mutator is free to take that at
653 // For scanning, that CPU time is
655 // assistDuration * procs * (utilization + idleUtilization)
657 // In this case, we *include* idle utilization, because that is additional CPU time that the
658 // the GC had available to it.
660 // In effect, idle GC time is sort of double-counted here, but it's very weird compared
661 // to other kinds of GC work, because of how fluid it is. Namely, because the mutator is
662 // *always* free to take it.
664 // So this calculation is really:
665 // (heapLive-trigger) / (assistDuration * procs * (1-utilization)) /
666 // (scanWork) / (assistDuration * procs * (utilization+idleUtilization)
668 // Note that because we only care about the ratio, assistDuration and procs cancel out.
669 scanWork := c.heapScanWork.Load() + c.stackScanWork.Load() + c.globalsScanWork.Load()
670 currentConsMark := (float64(c.heapLive-c.trigger) * (utilization + idleUtilization)) /
671 (float64(scanWork) * (1 - utilization))
673 // Update cons/mark controller.
674 // Period for this is 1 GC cycle.
675 oldConsMark := c.consMark
676 c.consMark = c.consMarkController.next(c.consMark, currentConsMark, 1.0)
678 if debug.gcpacertrace > 0 {
680 print("pacer: ", int(utilization*100), "% CPU (", int(gcGoalUtilization*100), " exp.) for ")
681 print(c.heapScanWork.Load(), "+", c.stackScanWork.Load(), "+", c.globalsScanWork.Load(), " B work (", c.lastHeapScan+c.stackScan+c.globalsScan, " B exp.) ")
682 print("in ", c.trigger, " B -> ", c.heapLive, " B (∆goal ", int64(c.heapLive)-int64(c.heapGoal), ", cons/mark ", oldConsMark, ")")
689 // !goexperiment.PacerRedesign below.
692 // Forced GC means this cycle didn't start at the
693 // trigger, so where it finished isn't good
694 // information about how to adjust the trigger.
695 // Just leave it where it is.
696 return c.triggerRatio
699 // Proportional response gain for the trigger controller. Must
700 // be in [0, 1]. Lower values smooth out transient effects but
701 // take longer to respond to phase changes. Higher values
702 // react to phase changes quickly, but are more affected by
703 // transient changes. Values near 1 may be unstable.
704 const triggerGain = 0.5
706 // Compute next cycle trigger ratio. First, this computes the
707 // "error" for this cycle; that is, how far off the trigger
708 // was from what it should have been, accounting for both heap
709 // growth and GC CPU utilization. We compute the actual heap
710 // growth during this cycle and scale that by how far off from
711 // the goal CPU utilization we were (to estimate the heap
712 // growth if we had the desired CPU utilization). The
713 // difference between this estimate and the GOGC-based goal
714 // heap growth is the error.
715 goalGrowthRatio := c.effectiveGrowthRatio()
716 actualGrowthRatio := float64(c.heapLive)/float64(c.heapMarked) - 1
717 triggerError := goalGrowthRatio - c.triggerRatio - utilization/gcGoalUtilization*(actualGrowthRatio-c.triggerRatio)
719 // Finally, we adjust the trigger for next time by this error,
720 // damped by the proportional gain.
721 triggerRatio := c.triggerRatio + triggerGain*triggerError
723 if debug.gcpacertrace > 0 {
724 // Print controller state in terms of the design
726 H_m_prev := c.heapMarked
727 h_t := c.triggerRatio
729 h_a := actualGrowthRatio
731 h_g := goalGrowthRatio
732 H_g := int64(float64(H_m_prev) * (1 + h_g))
734 u_g := gcGoalUtilization
735 W_a := c.heapScanWork.Load()
736 print("pacer: H_m_prev=", H_m_prev,
737 " h_t=", h_t, " H_T=", H_T,
738 " h_a=", h_a, " H_a=", H_a,
739 " h_g=", h_g, " H_g=", H_g,
740 " u_a=", u_a, " u_g=", u_g,
742 " goalΔ=", goalGrowthRatio-h_t,
743 " actualΔ=", h_a-h_t,
744 " u_a/u_g=", u_a/u_g,
751 // enlistWorker encourages another dedicated mark worker to start on
752 // another P if there are spare worker slots. It is used by putfull
753 // when more work is made available.
756 func (c *gcControllerState) enlistWorker() {
757 // If there are idle Ps, wake one so it will run an idle worker.
758 // NOTE: This is suspected of causing deadlocks. See golang.org/issue/19112.
760 // if atomic.Load(&sched.npidle) != 0 && atomic.Load(&sched.nmspinning) == 0 {
765 // There are no idle Ps. If we need more dedicated workers,
766 // try to preempt a running P so it will switch to a worker.
767 if c.dedicatedMarkWorkersNeeded <= 0 {
770 // Pick a random other P to preempt.
775 if gp == nil || gp.m == nil || gp.m.p == 0 {
778 myID := gp.m.p.ptr().id
779 for tries := 0; tries < 5; tries++ {
780 id := int32(fastrandn(uint32(gomaxprocs - 1)))
785 if p.status != _Prunning {
794 // findRunnableGCWorker returns a background mark worker for _p_ if it
795 // should be run. This must only be called when gcBlackenEnabled != 0.
796 func (c *gcControllerState) findRunnableGCWorker(_p_ *p) *g {
797 if gcBlackenEnabled == 0 {
798 throw("gcControllerState.findRunnable: blackening not enabled")
801 if !gcMarkWorkAvailable(_p_) {
802 // No work to be done right now. This can happen at
803 // the end of the mark phase when there are still
804 // assists tapering off. Don't bother running a worker
805 // now because it'll just return immediately.
809 // Grab a worker before we commit to running below.
810 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
812 // There is at least one worker per P, so normally there are
813 // enough workers to run on all Ps, if necessary. However, once
814 // a worker enters gcMarkDone it may park without rejoining the
815 // pool, thus freeing a P with no corresponding worker.
816 // gcMarkDone never depends on another worker doing work, so it
817 // is safe to simply do nothing here.
819 // If gcMarkDone bails out without completing the mark phase,
820 // it will always do so with queued global work. Thus, that P
821 // will be immediately eligible to re-run the worker G it was
822 // just using, ensuring work can complete.
826 decIfPositive := func(ptr *int64) bool {
828 v := atomic.Loadint64(ptr)
833 if atomic.Casint64(ptr, v, v-1) {
839 if decIfPositive(&c.dedicatedMarkWorkersNeeded) {
840 // This P is now dedicated to marking until the end of
841 // the concurrent mark phase.
842 _p_.gcMarkWorkerMode = gcMarkWorkerDedicatedMode
843 } else if c.fractionalUtilizationGoal == 0 {
844 // No need for fractional workers.
845 gcBgMarkWorkerPool.push(&node.node)
848 // Is this P behind on the fractional utilization
851 // This should be kept in sync with pollFractionalWorkerExit.
852 delta := nanotime() - c.markStartTime
853 if delta > 0 && float64(_p_.gcFractionalMarkTime)/float64(delta) > c.fractionalUtilizationGoal {
854 // Nope. No need to run a fractional worker.
855 gcBgMarkWorkerPool.push(&node.node)
858 // Run a fractional worker.
859 _p_.gcMarkWorkerMode = gcMarkWorkerFractionalMode
862 // Run the background mark worker.
864 casgstatus(gp, _Gwaiting, _Grunnable)
871 // resetLive sets up the controller state for the next mark phase after the end
872 // of the previous one. Must be called after endCycle and before commit, before
873 // the world is started.
875 // The world must be stopped.
876 func (c *gcControllerState) resetLive(bytesMarked uint64) {
877 c.heapMarked = bytesMarked
878 c.heapLive = bytesMarked
879 c.heapScan = uint64(c.heapScanWork.Load())
880 c.lastHeapScan = uint64(c.heapScanWork.Load())
882 // heapLive was updated, so emit a trace event.
888 // logWorkTime updates mark work accounting in the controller by a duration of
889 // work in nanoseconds.
891 // Safe to execute at any time.
892 func (c *gcControllerState) logWorkTime(mode gcMarkWorkerMode, duration int64) {
894 case gcMarkWorkerDedicatedMode:
895 atomic.Xaddint64(&c.dedicatedMarkTime, duration)
896 atomic.Xaddint64(&c.dedicatedMarkWorkersNeeded, 1)
897 case gcMarkWorkerFractionalMode:
898 atomic.Xaddint64(&c.fractionalMarkTime, duration)
899 case gcMarkWorkerIdleMode:
900 atomic.Xaddint64(&c.idleMarkTime, duration)
902 throw("logWorkTime: unknown mark worker mode")
906 func (c *gcControllerState) update(dHeapLive, dHeapScan int64) {
908 atomic.Xadd64(&gcController.heapLive, dHeapLive)
910 // gcController.heapLive changed.
914 // Only update heapScan in the new pacer redesign if we're not
915 // currently in a GC.
916 if !goexperiment.PacerRedesign || gcBlackenEnabled == 0 {
918 atomic.Xadd64(&gcController.heapScan, dHeapScan)
921 if gcBlackenEnabled != 0 {
922 // gcController.heapLive and heapScan changed.
927 func (c *gcControllerState) addScannableStack(pp *p, amount int64) {
929 atomic.Xadd64(&c.scannableStackSize, amount)
932 pp.scannableStackSizeDelta += amount
933 if pp.scannableStackSizeDelta >= scannableStackSizeSlack || pp.scannableStackSizeDelta <= -scannableStackSizeSlack {
934 atomic.Xadd64(&c.scannableStackSize, pp.scannableStackSizeDelta)
935 pp.scannableStackSizeDelta = 0
939 func (c *gcControllerState) addGlobals(amount int64) {
940 atomic.Xadd64(&c.globalsScan, amount)
943 // commit recomputes all pacing parameters from scratch, namely
944 // absolute trigger, the heap goal, mark pacing, and sweep pacing.
946 // If goexperiment.PacerRedesign is true, triggerRatio is ignored.
948 // This can be called any time. If GC is the in the middle of a
949 // concurrent phase, it will adjust the pacing of that phase.
951 // This depends on gcPercent, gcController.heapMarked, and
952 // gcController.heapLive. These must be up to date.
954 // mheap_.lock must be held or the world must be stopped.
955 func (c *gcControllerState) commit(triggerRatio float64) {
957 assertWorldStoppedOrLockHeld(&mheap_.lock)
960 if !goexperiment.PacerRedesign {
961 c.oldCommit(triggerRatio)
965 // Compute the next GC goal, which is when the allocated heap
966 // has grown by GOGC/100 over where it started the last cycle,
967 // plus additional runway for non-heap sources of GC work.
969 if gcPercent := c.gcPercent.Load(); gcPercent >= 0 {
970 goal = c.heapMarked + (c.heapMarked+atomic.Load64(&c.stackScan)+atomic.Load64(&c.globalsScan))*uint64(gcPercent)/100
973 // Don't trigger below the minimum heap size.
974 minTrigger := c.heapMinimum
976 // Concurrent sweep happens in the heap growth
977 // from gcController.heapLive to trigger, so ensure
978 // that concurrent sweep has some heap growth
979 // in which to perform sweeping before we
980 // start the next GC cycle.
981 sweepMin := atomic.Load64(&c.heapLive) + sweepMinHeapDistance
982 if sweepMin > minTrigger {
983 minTrigger = sweepMin
987 // If we let the trigger go too low, then if the application
988 // is allocating very rapidly we might end up in a situation
989 // where we're allocating black during a nearly always-on GC.
990 // The result of this is a growing heap and ultimately an
991 // increase in RSS. By capping us at a point >0, we're essentially
992 // saying that we're OK using more CPU during the GC to prevent
993 // this growth in RSS.
995 // The current constant was chosen empirically: given a sufficiently
996 // fast/scalable allocator with 48 Ps that could drive the trigger ratio
997 // to <0.05, this constant causes applications to retain the same peak
998 // RSS compared to not having this allocator.
999 if triggerBound := uint64(0.7*float64(goal-c.heapMarked)) + c.heapMarked; minTrigger < triggerBound {
1000 minTrigger = triggerBound
1003 // For small heaps, set the max trigger point at 95% of the heap goal.
1004 // This ensures we always have *some* headroom when the GC actually starts.
1005 // For larger heaps, set the max trigger point at the goal, minus the
1006 // minimum heap size.
1007 // This choice follows from the fact that the minimum heap size is chosen
1008 // to reflect the costs of a GC with no work to do. With a large heap but
1009 // very little scan work to perform, this gives us exactly as much runway
1010 // as we would need, in the worst case.
1011 maxRunway := uint64(0.95 * float64(goal-c.heapMarked))
1012 if largeHeapMaxRunway := goal - c.heapMinimum; goal > c.heapMinimum && maxRunway < largeHeapMaxRunway {
1013 maxRunway = largeHeapMaxRunway
1015 maxTrigger := maxRunway + c.heapMarked
1016 if maxTrigger < minTrigger {
1017 maxTrigger = minTrigger
1020 // Compute the trigger by using our estimate of the cons/mark ratio.
1022 // The idea is to take our expected scan work, and multiply it by
1023 // the cons/mark ratio to determine how long it'll take to complete
1024 // that scan work in terms of bytes allocated. This gives us our GC's
1027 // However, the cons/mark ratio is a ratio of rates per CPU-second, but
1028 // here we care about the relative rates for some division of CPU
1029 // resources among the mutator and the GC.
1031 // To summarize, we have B / cpu-ns, and we want B / ns. We get that
1032 // by multiplying by our desired division of CPU resources. We choose
1033 // to express CPU resources as GOMAPROCS*fraction. Note that because
1034 // we're working with a ratio here, we can omit the number of CPU cores,
1035 // because they'll appear in the numerator and denominator and cancel out.
1036 // As a result, this is basically just "weighing" the cons/mark ratio by
1037 // our desired division of resources.
1039 // Furthermore, by setting the trigger so that CPU resources are divided
1040 // this way, assuming that the cons/mark ratio is correct, we make that
1041 // division a reality.
1043 runway := uint64((c.consMark * (1 - gcGoalUtilization) / (gcGoalUtilization)) * float64(c.lastHeapScan+c.stackScan+c.globalsScan))
1045 trigger = minTrigger
1047 trigger = goal - runway
1049 if trigger < minTrigger {
1050 trigger = minTrigger
1052 if trigger > maxTrigger {
1053 trigger = maxTrigger
1059 // Commit to the trigger and goal.
1061 atomic.Store64(&c.heapGoal, goal)
1066 // Update mark pacing.
1067 if gcphase != _GCoff {
1072 // oldCommit sets the trigger ratio and updates everything
1073 // derived from it: the absolute trigger, the heap goal, mark pacing,
1074 // and sweep pacing.
1076 // This can be called any time. If GC is the in the middle of a
1077 // concurrent phase, it will adjust the pacing of that phase.
1079 // This depends on gcPercent, gcController.heapMarked, and
1080 // gcController.heapLive. These must be up to date.
1082 // For !goexperiment.PacerRedesign.
1083 func (c *gcControllerState) oldCommit(triggerRatio float64) {
1084 gcPercent := c.gcPercent.Load()
1086 // Compute the next GC goal, which is when the allocated heap
1087 // has grown by GOGC/100 over the heap marked by the last
1091 goal = c.heapMarked + c.heapMarked*uint64(gcPercent)/100
1094 // Set the trigger ratio, capped to reasonable bounds.
1096 scalingFactor := float64(gcPercent) / 100
1097 // Ensure there's always a little margin so that the
1098 // mutator assist ratio isn't infinity.
1099 maxTriggerRatio := 0.95 * scalingFactor
1100 if triggerRatio > maxTriggerRatio {
1101 triggerRatio = maxTriggerRatio
1104 // If we let triggerRatio go too low, then if the application
1105 // is allocating very rapidly we might end up in a situation
1106 // where we're allocating black during a nearly always-on GC.
1107 // The result of this is a growing heap and ultimately an
1108 // increase in RSS. By capping us at a point >0, we're essentially
1109 // saying that we're OK using more CPU during the GC to prevent
1110 // this growth in RSS.
1112 // The current constant was chosen empirically: given a sufficiently
1113 // fast/scalable allocator with 48 Ps that could drive the trigger ratio
1114 // to <0.05, this constant causes applications to retain the same peak
1115 // RSS compared to not having this allocator.
1116 minTriggerRatio := 0.6 * scalingFactor
1117 if triggerRatio < minTriggerRatio {
1118 triggerRatio = minTriggerRatio
1120 } else if triggerRatio < 0 {
1121 // gcPercent < 0, so just make sure we're not getting a negative
1122 // triggerRatio. This case isn't expected to happen in practice,
1123 // and doesn't really matter because if gcPercent < 0 then we won't
1124 // ever consume triggerRatio further on in this function, but let's
1125 // just be defensive here; the triggerRatio being negative is almost
1126 // certainly undesirable.
1129 c.triggerRatio = triggerRatio
1131 // Compute the absolute GC trigger from the trigger ratio.
1133 // We trigger the next GC cycle when the allocated heap has
1134 // grown by the trigger ratio over the marked heap size.
1135 trigger := ^uint64(0)
1137 trigger = uint64(float64(c.heapMarked) * (1 + triggerRatio))
1138 // Don't trigger below the minimum heap size.
1139 minTrigger := c.heapMinimum
1141 // Concurrent sweep happens in the heap growth
1142 // from gcController.heapLive to trigger, so ensure
1143 // that concurrent sweep has some heap growth
1144 // in which to perform sweeping before we
1145 // start the next GC cycle.
1146 sweepMin := atomic.Load64(&c.heapLive) + sweepMinHeapDistance
1147 if sweepMin > minTrigger {
1148 minTrigger = sweepMin
1151 if trigger < minTrigger {
1152 trigger = minTrigger
1154 if int64(trigger) < 0 {
1155 print("runtime: heapGoal=", c.heapGoal, " heapMarked=", c.heapMarked, " gcController.heapLive=", c.heapLive, " initialHeapLive=", work.initialHeapLive, "triggerRatio=", triggerRatio, " minTrigger=", minTrigger, "\n")
1156 throw("trigger underflow")
1159 // The trigger ratio is always less than GOGC/100, but
1160 // other bounds on the trigger may have raised it.
1161 // Push up the goal, too.
1166 // Commit to the trigger and goal.
1168 atomic.Store64(&c.heapGoal, goal)
1173 // Update mark pacing.
1174 if gcphase != _GCoff {
1179 // effectiveGrowthRatio returns the current effective heap growth
1180 // ratio (GOGC/100) based on heapMarked from the previous GC and
1181 // heapGoal for the current GC.
1183 // This may differ from gcPercent/100 because of various upper and
1184 // lower bounds on gcPercent. For example, if the heap is smaller than
1185 // heapMinimum, this can be higher than gcPercent/100.
1187 // mheap_.lock must be held or the world must be stopped.
1188 func (c *gcControllerState) effectiveGrowthRatio() float64 {
1190 assertWorldStoppedOrLockHeld(&mheap_.lock)
1193 egogc := float64(atomic.Load64(&c.heapGoal)-c.heapMarked) / float64(c.heapMarked)
1195 // Shouldn't happen, but just in case.
1201 // setGCPercent updates gcPercent and all related pacer state.
1202 // Returns the old value of gcPercent.
1204 // Calls gcControllerState.commit.
1206 // The world must be stopped, or mheap_.lock must be held.
1207 func (c *gcControllerState) setGCPercent(in int32) int32 {
1209 assertWorldStoppedOrLockHeld(&mheap_.lock)
1212 out := c.gcPercent.Load()
1216 c.heapMinimum = defaultHeapMinimum * uint64(in) / 100
1217 c.gcPercent.Store(in)
1218 // Update pacing in response to gcPercent change.
1219 c.commit(c.triggerRatio)
1224 //go:linkname setGCPercent runtime/debug.setGCPercent
1225 func setGCPercent(in int32) (out int32) {
1226 // Run on the system stack since we grab the heap lock.
1227 systemstack(func() {
1229 out = gcController.setGCPercent(in)
1230 gcPaceSweeper(gcController.trigger)
1231 gcPaceScavenger(gcController.heapGoal, gcController.lastHeapGoal)
1232 unlock(&mheap_.lock)
1235 // If we just disabled GC, wait for any concurrent GC mark to
1236 // finish so we always return with no GC running.
1238 gcWaitOnMark(atomic.Load(&work.cycles))
1244 func readGOGC() int32 {
1245 p := gogetenv("GOGC")
1249 if n, ok := atoi32(p); ok {
1255 type piController struct {
1256 kp float64 // Proportional constant.
1257 ti float64 // Integral time constant.
1258 tt float64 // Reset time.
1260 min, max float64 // Output boundaries.
1262 // PI controller state.
1264 errIntegral float64 // Integral of the error from t=0 to now.
1267 func (c *piController) next(input, setpoint, period float64) float64 {
1268 // Compute the raw output value.
1269 prop := c.kp * (setpoint - input)
1270 rawOutput := prop + c.errIntegral
1272 // Clamp rawOutput into output.
1276 } else if output > c.max {
1280 // Update the controller's state.
1281 if c.ti != 0 && c.tt != 0 {
1282 c.errIntegral += (c.kp*period/c.ti)*(setpoint-input) + (period/c.tt)*(output-rawOutput)