This moves all of GC initialization, sweep termination, and the
transition to concurrent marking in to the off->mark transition
function. This means it's now handled on the goroutine that detected
the state exit condition.
As a result, malloc no longer needs to Gosched() at the beginning of
the GC cycle to prevent over-allocation while the GC is starting up
because it will now *help* the GC to start up. The Gosched hack is
still necessary during GC shutdown (this is easy to test by enabling
gctrace and hitting Ctrl-S to block the gctrace output).
At this point, the GC coordinator still handles later phases. This
requires a small tweak to how we start the GC coordinator. Currently,
starting the GC coordinator is best-effort and may fail if the
coordinator is about to park from the previous cycle but hasn't yet.
We fix this by replacing the park/ready to wake up the coordinator
with a semaphore. This is temporary since the coordinator will be
going away in a few commits.
Updates #11970.
Change-Id: I2c6a11c91e72dfbc59c2d8e7c66146dee9a444fe
Reviewed-on: https://go-review.googlesource.com/16357
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This moves concurrent sweep termination from the coordinator to the
off->mark transition. This allows it to be performed by all Gs
attempting to start the GC.
Updates #11970.
Change-Id: I24428e8599a759398c2ef7ec996ba755a448f947
Reviewed-on: https://go-review.googlesource.com/16356
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This begins the conversion of the centralized GC coordinator to a
decentralized state machine by introducing the internal API that
triggers the first state transition from _GCoff to _GCmark (or
_GCmarktermination).
This change introduces the transition lock, the off->mark transition
condition (which is very similar to shouldtriggergc()), and the
general structure of a state transition. Since we're doing this
conversion in stages, it then falls back to the GC coordinator to
actually execute the cycle. We'll start moving logic out of the GC
coordinator and in to transition functions next.
This fixes a minor bug in gcstoptheworld debug mode where passing the
heap trigger once could trigger multiple STW GCs.
Updates #11970.
Change-Id: I964087dd190a639eb5766398f8e1bbf8b352902f
Reviewed-on: https://go-review.googlesource.com/16355
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
For historical reasons we currently do a lot of the concurrent mark
setup on the system stack. In fact, at this point the one and only
thing that needs to happen on the system stack is the start-the-world.
Clean up this code by lifting everything other than the
start-the-world off the system stack.
The diff for this change looks large, but the only code change is to
narrow the systemstack call. Everything else is re-indentation.
Change-Id: I1e03b8afc759fad726f2397b05a17d183c2713ce
Reviewed-on: https://go-review.googlesource.com/16354
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
We're about to split func gc across several functions, so lift the
local variables it uses for tracking statistics and state across the
cycle into the global "work" variable.
Change-Id: Ie955f2f1758c7f5a5543ea1f3f33b222bc4b1d37
Reviewed-on: https://go-review.googlesource.com/16353
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Currently dedicated mark workers participate in the getfull barrier
during concurrent mark. However, the getfull barrier wasn't designed
for concurrent work and this causes no end of headaches.
In the concurrent setting, participants come and go. This makes mark
completion susceptible to live-lock: since dedicated workers are only
periodically polling for completion, it's possible for the program to
be in some transient worker each time one of the dedicated workers
wakes up to check if it can exit the getfull barrier. It also
complicates reasoning about the system because dedicated workers
participate directly in the getfull barrier, but transient workers
must instead use trygetfull because they have exit conditions that
aren't captured by getfull (e.g., fractional workers exit when
preempted). The complexity of implementing these exit conditions
contributed to #11677. Furthermore, the getfull barrier is inefficient
because we could be running user code instead of spinning on a P. In
effect, we're dedicating 25% of the CPU to marking even if that means
we have to spin to make that 25%. It also causes issues on Windows
because we can't actually sleep for 100µs (#8687).
Fix this by making dedicated workers no longer participate in the
getfull barrier. Instead, dedicated workers simply return to the
scheduler when they fail to get more work, regardless of what others
workers are doing, and the scheduler only starts new dedicated workers
if there's work available. Everything that needs to be handled by this
barrier is already handled by detection of mark completion.
This makes the system much more symmetric because all workers and
assists now use trygetfull during concurrent mark. It also loosens the
25% CPU target so that we can give some of that 25% back to user code
if there isn't enough work to keep the mark worker busy. And it
eliminates the problematic 100µs sleep on Windows during concurrent
mark (though not during mark termination).
The downside of this is that if we hit a bottleneck in the heap graph
that then expands back out, the system may shut down dedicated workers
and take a while to start them back up. We'll address this in the next
commit.
Updates #12041 and #8687.
No effect on the go1 benchmarks. This slows down the garbage benchmark
by 9%, but we'll more than make it up in the next commit.
name old time/op new time/op delta
XBenchGarbage-12 5.80ms ± 2% 6.32ms ± 4% +9.03% (p=0.000 n=20+20)
Change-Id: I65100a9ba005a8b5cf97940798918672ea9dd09b
Reviewed-on: https://go-review.googlesource.com/16297
Reviewed-by: Rick Hudson <rlh@golang.org>
GC assists must block until the assist can be satisfied (either
through stealing credit or doing work) or the GC cycle ends.
Currently, this is implemented as a retry loop with a 100 µs delay.
This obviously isn't ideal, as it wastes CPU and delays mutator
execution. It also has the somewhat peculiar downside that sleeping a
G requires allocation, and this requires working around recursive
allocation.
Replace this timed delay with a proper scheduling queue. When an
assist can't be satisfied immediately, it adds the allocating G to a
queue and parks it. Any time background scan credit is flushed, it
consults this queue, directly satisfies the debt of queued assists,
and wakes up satisfied assists before flushing any remaining credit to
the background credit pool.
No effect on the go1 benchmarks. Slightly speeds up the garbage
benchmark.
name old time/op new time/op delta
XBenchGarbage-12 5.81ms ± 1% 5.72ms ± 4% -1.65% (p=0.011 n=20+20)
Updates #12041.
Change-Id: I8ee3b6274dd097b12b10a8030796a958a4b0e7b7
Reviewed-on: https://go-review.googlesource.com/15890
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Currently the concurrent root scan is performed in its entirety by the
GC coordinator before entering concurrent mark (which enables GC
workers). This scan is done sequentially, which can prolong the scan
phase, delay the mark phase, and means that the scan phase does not
obey the 25% CPU goal. Furthermore, there's no need to complete the
root scan before starting marking (in fact, we already allow GC
assists to happen during the scan phase), so this acts as an
unnecessary barrier between root scanning and marking.
This change shifts the root scan work out of the GC coordinator and in
to the GC workers. The coordinator simply sets up the scan state and
enqueues the right number of root scan jobs. The GC workers then drain
the root scan jobs prior to draining heap scan jobs.
This parallelizes the root scan process, makes it obey the 25% CPU
goal, and effectively eliminates root scanning as an isolated phase,
allowing the system to smoothly transition from root scanning to heap
marking. This also eliminates a major non-STW responsibility of the GC
coordinator, which will make it easier to switch to a decentralized
state machine. Finally, it puts us in a good position to perform root
scanning in assists as well, which will help satisfy assists at the
beginning of the GC cycle.
This is mostly straightforward. One tricky aspect is that we have to
deal with preemption deadlock: where two non-preemptible gorountines
are trying to preempt each other to perform a stack scan. Given the
context where this happens, the only instance of this is two
background workers trying to scan each other. We avoid this by simply
not scanning the stacks of background workers during the concurrent
phase; this is safe because we'll scan them during mark termination
(and their stacks are *very* small and should not contain any new
pointers).
This change also switches the root marking during mark termination to
use the same gcDrain-based code path as concurrent mark. This
shouldn't affect performance because STW root marking was already
parallel and tasks switched to heap marking immediately when no more
root marking tasks were available. However, it simplifies the code and
unifies these code paths.
This has negligible effect on the go1 benchmarks. It slightly slows
down the garbage benchmark, possibly by making GC run slightly more
frequently.
name old time/op new time/op delta
XBenchGarbage-12 5.10ms ± 1% 5.24ms ± 1% +2.87% (p=0.000 n=18+18)
name old time/op new time/op delta
BinaryTree17-12 3.25s ± 3% 3.20s ± 5% -1.57% (p=0.013 n=20+20)
Fannkuch11-12 2.45s ± 1% 2.46s ± 1% +0.38% (p=0.019 n=20+18)
FmtFprintfEmpty-12 49.7ns ± 3% 49.9ns ± 4% ~ (p=0.851 n=19+20)
FmtFprintfString-12 170ns ± 2% 170ns ± 1% ~ (p=0.775 n=20+19)
FmtFprintfInt-12 161ns ± 1% 160ns ± 1% -0.78% (p=0.000 n=19+18)
FmtFprintfIntInt-12 267ns ± 1% 270ns ± 1% +1.04% (p=0.000 n=19+19)
FmtFprintfPrefixedInt-12 238ns ± 2% 238ns ± 1% ~ (p=0.133 n=18+19)
FmtFprintfFloat-12 311ns ± 1% 310ns ± 2% -0.35% (p=0.023 n=20+19)
FmtManyArgs-12 1.08µs ± 1% 1.06µs ± 1% -2.31% (p=0.000 n=20+20)
GobDecode-12 8.65ms ± 1% 8.63ms ± 1% ~ (p=0.377 n=18+20)
GobEncode-12 6.49ms ± 1% 6.52ms ± 1% +0.37% (p=0.015 n=20+20)
Gzip-12 319ms ± 3% 318ms ± 1% ~ (p=0.975 n=19+17)
Gunzip-12 41.9ms ± 1% 42.1ms ± 2% +0.65% (p=0.004 n=19+20)
HTTPClientServer-12 61.7µs ± 1% 62.6µs ± 1% +1.40% (p=0.000 n=18+20)
JSONEncode-12 16.8ms ± 1% 16.9ms ± 1% ~ (p=0.239 n=20+18)
JSONDecode-12 58.4ms ± 1% 60.7ms ± 1% +3.85% (p=0.000 n=19+20)
Mandelbrot200-12 3.86ms ± 0% 3.86ms ± 1% ~ (p=0.092 n=18+19)
GoParse-12 3.75ms ± 2% 3.75ms ± 2% ~ (p=0.708 n=19+20)
RegexpMatchEasy0_32-12 100ns ± 1% 100ns ± 2% +0.60% (p=0.010 n=17+20)
RegexpMatchEasy0_1K-12 341ns ± 1% 342ns ± 2% ~ (p=0.203 n=20+19)
RegexpMatchEasy1_32-12 82.5ns ± 2% 83.2ns ± 2% +0.83% (p=0.007 n=19+19)
RegexpMatchEasy1_1K-12 495ns ± 1% 495ns ± 2% ~ (p=0.970 n=19+18)
RegexpMatchMedium_32-12 130ns ± 2% 130ns ± 2% +0.59% (p=0.039 n=19+20)
RegexpMatchMedium_1K-12 39.2µs ± 1% 39.3µs ± 1% ~ (p=0.214 n=18+18)
RegexpMatchHard_32-12 2.03µs ± 2% 2.02µs ± 1% ~ (p=0.166 n=18+19)
RegexpMatchHard_1K-12 61.0µs ± 1% 60.9µs ± 1% ~ (p=0.169 n=20+18)
Revcomp-12 533ms ± 1% 535ms ± 1% ~ (p=0.071 n=19+17)
Template-12 68.1ms ± 2% 73.0ms ± 1% +7.26% (p=0.000 n=19+20)
TimeParse-12 355ns ± 2% 356ns ± 2% ~ (p=0.530 n=19+20)
TimeFormat-12 357ns ± 2% 347ns ± 1% -2.59% (p=0.000 n=20+19)
[Geo mean] 62.1µs 62.3µs +0.31%
name old speed new speed delta
GobDecode-12 88.7MB/s ± 1% 88.9MB/s ± 1% ~ (p=0.377 n=18+20)
GobEncode-12 118MB/s ± 1% 118MB/s ± 1% -0.37% (p=0.015 n=20+20)
Gzip-12 60.9MB/s ± 3% 60.9MB/s ± 1% ~ (p=0.944 n=19+17)
Gunzip-12 464MB/s ± 1% 461MB/s ± 2% -0.64% (p=0.004 n=19+20)
JSONEncode-12 115MB/s ± 1% 115MB/s ± 1% ~ (p=0.236 n=20+18)
JSONDecode-12 33.2MB/s ± 1% 32.0MB/s ± 1% -3.71% (p=0.000 n=19+20)
GoParse-12 15.5MB/s ± 2% 15.5MB/s ± 2% ~ (p=0.702 n=19+20)
RegexpMatchEasy0_32-12 320MB/s ± 1% 318MB/s ± 2% ~ (p=0.094 n=18+20)
RegexpMatchEasy0_1K-12 3.00GB/s ± 1% 2.99GB/s ± 1% ~ (p=0.194 n=20+19)
RegexpMatchEasy1_32-12 388MB/s ± 2% 385MB/s ± 2% -0.83% (p=0.008 n=19+19)
RegexpMatchEasy1_1K-12 2.07GB/s ± 1% 2.07GB/s ± 1% ~ (p=0.964 n=19+18)
RegexpMatchMedium_32-12 7.68MB/s ± 1% 7.64MB/s ± 2% -0.57% (p=0.020 n=19+20)
RegexpMatchMedium_1K-12 26.1MB/s ± 1% 26.1MB/s ± 1% ~ (p=0.211 n=18+18)
RegexpMatchHard_32-12 15.8MB/s ± 1% 15.8MB/s ± 1% ~ (p=0.180 n=18+19)
RegexpMatchHard_1K-12 16.8MB/s ± 1% 16.8MB/s ± 2% ~ (p=0.236 n=20+19)
Revcomp-12 477MB/s ± 1% 475MB/s ± 1% ~ (p=0.071 n=19+17)
Template-12 28.5MB/s ± 2% 26.6MB/s ± 1% -6.77% (p=0.000 n=19+20)
[Geo mean] 100MB/s 99.0MB/s -0.82%
Change-Id: I875bf6ceb306d1ee2f470cabf88aa6ede27c47a0
Reviewed-on: https://go-review.googlesource.com/16059
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
We already have gcMarkWorkAvailable, but the check for GC mark work is
open-coded in several places. Generalize gcMarkWorkAvailable slightly
and replace these open-coded checks with calls to gcMarkWorkAvailable.
In addition to cleaning up the code, this puts us in a better position
to make this check slightly more complicated.
Change-Id: I1b29883300ecd82a1bf6be193e9b4ee96582a860
Reviewed-on: https://go-review.googlesource.com/16058
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
The ragged barrier after entering the concurrent mark phase is
vestigial. This used to be the point where we enabled write barriers,
so it was necessary to synchronize all Ps to ensure write barriers
were enabled before any marking occurred. However, we've long since
switched to enabling write barriers during the concurrent scan phase,
so the start-the-world at the beginning of the concurrent scan phase
ensures that all Ps have enabled the write barrier.
Hence, we can eliminate the old "install write barrier" phase.
Fixes#11971.
Change-Id: I8cdcb84b5525cef19927d51ea11ba0a4db991ea8
Reviewed-on: https://go-review.googlesource.com/16044
Reviewed-by: Rick Hudson <rlh@golang.org>
These functions are always called together and perform logically
related state resets, so combine them in to just gcResetMarkState.
Fixes#11427.
Change-Id: I06c17ef65f66186494887a767b3993126955b5fe
Reviewed-on: https://go-review.googlesource.com/16041
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently gcResetGState is called by func gcscan_m for concurrent GC
and directly by func gc for STW GC. Simplify this by consolidating
these two calls in to one call by func gc above where it splits for
concurrent and STW GC.
As a consequence, gcResetGState and gcResetMarkState are always called
together, so the next commit will consolidate these.
Change-Id: Ib62d404c7b32b28f7d3080d26ecf3966cbc4aca0
Reviewed-on: https://go-review.googlesource.com/16040
Reviewed-by: Rick Hudson <rlh@golang.org>
This work queue is no longer used (there are many reads of
work.partial, but the only write is in putpartial, which is never
called).
Fixes#11922.
Change-Id: I08b76c0c02a0867a9cdcb94783e1f7629d44249a
Reviewed-on: https://go-review.googlesource.com/15892
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently, when the mutator allocates, the runtime first allocates the
memory and then, if that G has done "enough" allocation, the runtime
checks whether the G has assist debt to pay off and, if so, pays it
off. This approach leads to under-assisting, where a G can allocate a
large region (or many small regions) before paying for it, or can even
exit with outstanding debt.
This commit flips this around so that a G always acquires enough
credit for an allocation before it can perform that allocation. We
continue to amortize the cost of assists by requiring that they
over-assist when triggered to build up credit for many allocations.
Fixes#11967.
Change-Id: Idac9f11133b328535667674d837be72c23ebd899
Reviewed-on: https://go-review.googlesource.com/15409
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
Currently we track the per-G GC assist balance as two monotonically
increasing values: the bytes allocated by the G this cycle (gcalloc)
and the scan work performed by the G this cycle (gcscanwork). The
assist balance is hence assistRatio*gcalloc - gcscanwork.
This works, but has two important downsides:
1) It requires floating-point math to figure out if a G is in debt or
not. This makes it inappropriate to check for assist debt in the
hot path of mallocgc, so we only do this when a G allocates a new
span. As a result, Gs can operate "in the red", leading to
under-assist and extended GC cycle length.
2) Revising the assist ratio during a GC cycle can lead to an "assist
burst". If you think of plotting the scan work performed versus
heaps size, the assist ratio controls the slope of this line.
However, in the current system, the target line always passes
through 0 at the heap size that triggered GC, so if the runtime
increases the assist ratio, there has to be a potentially large
assist to jump from the current amount of scan work up to the new
target scan work for the current heap size.
This commit replaces this approach with directly tracking the GC
assist balance in terms of allocation credit bytes. Allocating N bytes
simply decreases this by N and assisting raises it by the amount of
scan work performed divided by the assist ratio (to get back to
bytes).
This will make it cheap to figure out if a G is in debt, which will
let us efficiently check if an assist is necessary *before* performing
an allocation and hence keep Gs "in the black".
This also fixes assist bursts because the assist ratio is now in terms
of *remaining* work, rather than work from the beginning of the GC
cycle. Hence, the plot of scan work versus heap size becomes
continuous: we can revise the slope, but this slope always starts from
where we are right now, rather than where we were at the beginning of
the cycle.
Change-Id: Ia821c5f07f8a433e8da7f195b52adfedd58bdf2c
Reviewed-on: https://go-review.googlesource.com/15408
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently we ensure a minimum heap distance of 1MB when computing the
assist ratio. Rather than enforcing this minimum on the heap distance,
it makes more sense to enforce that the heap goal itself is at least
1MB over the live heap size at the beginning of GC. Currently the two
approaches are semantically equivalent, but this will let us switch to
basing the assist ratio on current heap distance rather than the
initial heap distance, since we can't enforce this minimum on the
current heap distance (the GC may never finish because the goal posts
will always be 1MB away).
Change-Id: I0027b1c26a41a0152b01e5b67bdb1140d43ee903
Reviewed-on: https://go-review.googlesource.com/15604
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently, gcController.scanWork is updated as lazily as possible
since it is only read at the end of the GC cycle. We're about to read
it during the GC cycle to improve the assist ratio revisions, so
modify gcDrain* to regularly flush to gcController.scanWork in much
the same way as we regularly flush to gcController.bgScanCredit.
One consequence of this is that it's difficult to keep gcw.scanWork
monotonic, so we give up on that and simply return the amount of scan
work done by gcDrainN rather than calculating it in the caller.
Change-Id: I7b50acdc39602f843eed0b5c6d2dacd7e762b81d
Reviewed-on: https://go-review.googlesource.com/15407
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently callers of gcDrain control whether it flushes scan work
credit to gcController.bgScanCredit by passing a value other than -1
for the flush threshold. Shortly we're going to make this always flush
scan work to gcController.scanWork and optionally also flush scan work
to gcController.bgScanCredit. This will be much easier if the flush
threshold is simply a constant (which it is in practice) and callers
merely control whether or not the flush includes the background
credit. Hence, replace the flush threshold argument with a flag.
Change-Id: Ia27db17de8a3f1e462a5d7137d4b5dc72f99a04e
Reviewed-on: https://go-review.googlesource.com/15406
Reviewed-by: Rick Hudson <rlh@golang.org>
These functions were nearly identical. Consolidate them by adding a
flags argument. In addition to cleaning up this code, this makes
further changes that affect both functions easier.
Change-Id: I6ec5c947603bbbd3ff4040113b2fbc240e99745f
Reviewed-on: https://go-review.googlesource.com/15405
Reviewed-by: Rick Hudson <rlh@golang.org>
The comment for assistRatio claimed it to be the reciprocal of what it
actually is.
Change-Id: If7f9bb853d75d0097facff3aa6704b224d9108b8
Reviewed-on: https://go-review.googlesource.com/15402
Reviewed-by: Russ Cox <rsc@golang.org>
In general, finishsweep_m must block until any spans that are
concurrently being swept have been swept. It accomplishes this by
looping over all spans, which, as in the previous commit, takes
~1ms/heap GB. Unfortunately, we do this during the STW sweep
termination phase, so multi-gigabyte heaps can push our STW time past
10ms.
However, there's no need to do this wait if the world is stopped
because, in effect, stopping the world already had to wait for
anything that was sweeping (and if it didn't, the wait in
finishsweep_m would deadlock). Hence, we can simply skip this loop if
the world is stopped, such as during sweep termination. In fact,
currently all calls to finishsweep_m are STW, but this hasn't always
been the case and may not be the case in the future, so we keep the
logic around.
For 24GB heaps, this reduces max pause time by 75% relative to tip and
by 90% relative to Go 1.5. Notably, all pauses are now well under
10ms. Here are the results for the garbage benchmark:
------------- max pause ------------
Heap Procs after change before change 1.5.1
24GB 12 3.8ms 16ms 37ms
24GB 4 3.7ms 16ms 37ms
4GB 4 3.7ms 3ms 6.9ms
In the 4GB/4P case, it seems the "before change" run got lucky: the
max went up, but the 99%ile pause time went down from 3ms to 2.04ms.
Change-Id: Ica22189559f231d408ef2815019c9dbb5f38bf31
Reviewed-on: https://go-review.googlesource.com/15071
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
In order to compute the sweep ratio, the runtime needs to know how
many pages belong to spans in state _MSpanInUse. Currently it finds
this out by looping over all spans during mark termination. However,
this takes ~1ms/heap GB, so multi-gigabyte heaps can quickly push our
STW time past 10ms.
Replace the loop with an actively maintained count of in-use pages.
For multi-gigabyte heaps, this reduces max mark termination pause time
by 75%–90% relative to tip and by 85%–95% relative to Go 1.5.1. This
shifts the longest pause time for large heaps to the sweep termination
phase, so it only slightly decreases max pause time, though it roughly
halves mean pause time. Here are the results for the garbage
benchmark:
---- max mark termination pause ----
Heap Procs after change before change 1.5.1
24GB 12 1.9ms 18ms 37ms
24GB 4 3.7ms 18ms 37ms
4GB 4 920µs 3.8ms 6.9ms
Fixes#11484.
Change-Id: Ia2d28bb8a1e4f1c3b8ebf79fb203f12b9bf114ac
Reviewed-on: https://go-review.googlesource.com/15070
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This reduces pause time by ~25% relative to tip and by ~50% relative
to Go 1.5.1.
Currently one of the steps of STW mark termination is to loop (in
parallel) over all spans to find objects with finalizers in order to
mark all objects reachable from these objects and to treat the
finalizer special as a root. Unfortunately, even if there are no
finalizers at all, this loop takes roughly 1 ms/heap GB/core, so
multi-gigabyte heaps can quickly push our STW time past 10ms.
Fix this by moving this scan from mark termination to concurrent scan,
where it can run in parallel with mutators. The loop itself could also
be optimized, but this cost is small compared to concurrent marking.
Making this scan concurrent introduces two complications:
1) The scan currently walks the specials list of each span without
locking it, which is safe only with the world stopped. We fix this by
speculatively checking if a span has any specials (the vast majority
won't) and then locking the specials list only if there are specials
to check.
2) An object can have a finalizer set after concurrent scan, in which
case it won't have been marked appropriately by concurrent scan. If
the finalizer is a closure and is only reachable from the special, it
could be swept before it is run. Likewise, if the object is not marked
yet when the finalizer is set and then becomes unreachable before it
is marked, other objects reachable only from it may be swept before
the finalizer function is run. We fix this issue by making
addfinalizer ensure the same marking invariants as markroot does.
For multi-gigabyte heaps, this reduces max pause time by 20%–30%
relative to tip (depending on GOMAXPROCS) and by ~50% relative to Go
1.5.1 (where this loop was neither concurrent nor parallel). Here are
the results for the garbage benchmark:
---------------- max pause ----------------
Heap Procs Concurrent scan STW parallel scan 1.5.1
24GB 12 18ms 23ms 37ms
24GB 4 18ms 25ms 37ms
4GB 4 3.8ms 4.9ms 6.9ms
In all cases, 95%ile pause time is similar to the max pause time. This
also improves mean STW time by 10%–30%.
Fixes#11485.
Change-Id: I9359d8c3d120a51d23d924b52bf853a1299b1dfd
Reviewed-on: https://go-review.googlesource.com/14982
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Currently, the GC modes constants are untyped and functions pass them
around as ints. Clean this up by introducing a proper type for these
constant.
Change-Id: Ibc022447bdfa203644921fbb548312d7e2272e8d
Reviewed-on: https://go-review.googlesource.com/14981
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This puts the _Root* indexes in a more friendly order and tweaks
markrootSpans to use a for-range loop instead of its own indexing.
Change-Id: I2c18d55c9a673ea396b6424d51ef4997a1a74825
Reviewed-on: https://go-review.googlesource.com/14548
Reviewed-by: Rick Hudson <rlh@golang.org>
Marking of span roots can represent a significant fraction of the time
spent in mark termination. Simply traversing the span list takes about
1ms per GB of heap and if there are a large number of finalizers (for
example, for network connections), it may take much longer.
Improve the situation by splitting the span scan into 128 subtasks
that can be executed in parallel and load balanced by the markroots
parallel for. This lets the GC balance this job across the Ps.
A better solution is to do this during concurrent mark, or to improve
it algorithmically, but this is a simple change with a lot of bang for
the buck.
This was suggested by Rhys Hiltner.
Updates #11485.
Change-Id: I8b281adf0ba827064e154a1b6cc32d4d8031c03c
Reviewed-on: https://go-review.googlesource.com/13112
Reviewed-by: Keith Randall <khr@golang.org>
Currently the stack barrier code is mixed in with the mark and scan
code. Move all of the stack barrier related functions and variables to
a new dedicated source file. There are no code modifications.
Change-Id: I604603045465ef8573b9f88915d28ab6b5910903
Reviewed-on: https://go-review.googlesource.com/14050
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently enabling the debugging mode where stack barriers are
installed at every frame requires recompiling the runtime. However,
this is potentially useful for field debugging and for runtime tests,
so make this mode a GODEBUG.
Updates #12238.
Change-Id: I6fb128f598b19568ae723a612e099c0ed96917f5
Reviewed-on: https://go-review.googlesource.com/13947
Reviewed-by: Russ Cox <rsc@golang.org>
88e945f introduced a non-speculative double check of the heap trigger
before actually starting a concurrent GC. This was necessary to fix a
race for heap-triggered GC, but broke sysmon-triggered periodic GC,
since the heap check will of course fail for periodically triggered
GC.
Fix this by telling startGC whether or not this GC was triggered by
heap size or a timer and only doing the heap size double check for GCs
triggered by heap size.
Fixes#12026.
Change-Id: I7c3f6ec364545c36d619f2b4b3bf3b758e3bcbd6
Reviewed-on: https://go-review.googlesource.com/13168
Reviewed-by: Russ Cox <rsc@golang.org>
When commit 510fd13 enabled assists during the scan phase, it failed
to also update the code in the GC controller that computed the assist
CPU utilization and adjusted the trigger based on it. Fix that code so
it uses the start of the scan phase as the wall-clock time when
assists were enabled rather than the start of the mark phase.
Change-Id: I05013734b4448c3e2c730dc7b0b5ee28c86ed8cf
Reviewed-on: https://go-review.googlesource.com/13048
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
At the start of a GC cycle, the garbage collector computes the assist
ratio based on the total scannable heap size. This was intended to be
conservative; after all, this assumes the entire heap may be reachable
and hence needs to be scanned. But it only assumes that the *current*
entire heap may be reachable. It fails to account for heap allocated
during the GC cycle. If the trigger ratio is very low (near zero), and
most of the heap is reachable when GC starts (which is likely if the
trigger ratio is near zero), then it's possible for the mutator to
create new, reachable heap fast enough that the assists won't keep up
based on the assist ratio computed at the beginning of the cycle. As a
result, the heap can grow beyond the heap goal (by hundreds of megs in
stress tests like in issue #11911).
We already have some vestigial logic for dealing with situations like
this; it just doesn't run often enough. Currently, every 10 ms during
the GC cycle, the GC revises the assist ratio. This was put in before
we switched to a conservative assist ratio (when we really were using
estimates of scannable heap), and it turns out to be exactly what we
need now. However, every 10 ms is far too infrequent for a rapidly
allocating mutator.
This commit reuses this logic, but replaces the 10 ms timer with
revising the assist ratio every time the heap is locked, which
coincides precisely with when the statistics used to compute the
assist ratio are updated.
Fixes#11911.
Change-Id: I377b231ab064946228378fa10422a46d1b50f4c5
Reviewed-on: https://go-review.googlesource.com/13047
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
This was useful in debugging the mutator assist behavior for #11911,
and it fits with the other gcpacertrace output.
Change-Id: I1e25590bb4098223a160de796578bd11086309c7
Reviewed-on: https://go-review.googlesource.com/13046
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
Proportional concurrent sweep is currently based on a ratio of spans
to be swept per bytes of object allocation. However, proportional
sweeping is performed during span allocation, not object allocation,
in order to minimize contention and overhead. Since objects are
allocated from spans after those spans are allocated, the system tends
to operate in debt, which means when the next GC cycle starts, there
is often sweep debt remaining, so GC has to finish the sweep, which
delays the start of the cycle and delays enabling mutator assists.
For example, it's quite likely that many Ps will simultaneously refill
their span caches immediately after a GC cycle (because GC flushes the
span caches), but at this point, there has been very little object
allocation since the end of GC, so very little sweeping is done. The
Ps then allocate objects from these cached spans, which drives up the
bytes of object allocation, but since these allocations are coming
from cached spans, nothing considers whether more sweeping has to
happen. If the sweep ratio is high enough (which can happen if the
next GC trigger is very close to the retained heap size), this can
easily represent a sweep debt of thousands of pages.
Fix this by making proportional sweep proportional to the number of
bytes of spans allocated, rather than the number of bytes of objects
allocated. Prior to allocating a span, both the small object path and
the large object path ensure credit for allocating that span, so the
system operates in the black, rather than in the red.
Combined with the previous commit, this should eliminate all sweeping
from GC start up. On the stress test in issue #11911, this reduces the
time spent sweeping during GC (and delaying start up) by several
orders of magnitude:
mean 99%ile max
pre fix 1 ms 11 ms 144 ms
post fix 270 ns 735 ns 916 ns
Updates #11911.
Change-Id: I89223712883954c9d6ec2a7a51ecb97172097df3
Reviewed-on: https://go-review.googlesource.com/13044
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
Currently it's possible for the next_gc heap size trigger computed for
the next GC cycle to be less than the current allocated heap size.
This means the next cycle will start immediately, which means there's
no time to perform the concurrent sweep between GC cycles. This places
responsibility for finishing the sweep on GC itself, which delays GC
start-up and hence delays mutator assist.
Fix this by ensuring that next_gc is always at least a little higher
than the allocated heap size, so we won't trigger the next cycle
instantly.
Updates #11911.
Change-Id: I74f0b887bf187518d5fedffc7989817cbcf30592
Reviewed-on: https://go-review.googlesource.com/13043
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
Currently allocation checks the GC trigger speculatively during
allocation and then triggers the GC without rechecking. As a result,
it's possible for G 1 and G 2 to detect the trigger simultaneously,
both enter startGC, G 1 actually starts GC while G 2 gets preempted
until after the whole GC cycle, then G 2 immediately starts another GC
cycle even though the heap is now well under the trigger.
Fix this by re-checking the GC trigger non-speculatively just before
actually kicking off a new GC cycle.
This contributes to #11911 because when this happens, we definitely
don't finish the background sweep before starting the next GC cycle,
which can significantly delay the start of concurrent scan.
Change-Id: I560ab79ba5684ba435084410a9765d28f5745976
Reviewed-on: https://go-review.googlesource.com/13025
Reviewed-by: Russ Cox <rsc@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Dangling pointer error. Unlikely to trigger in practice, but still.
Found by running GODEBUG=efence=1 GOGC=1 trace.test.
Change-Id: Ice474dedcf62dd33ab77526287a023ba3b166db9
Reviewed-on: https://go-review.googlesource.com/12991
Reviewed-by: Austin Clements <austin@google.com>
This adds a GCCPUFraction field to MemStats that reports the
cumulative fraction of the program's execution time spent in the
garbage collector. This is equivalent to the utilization percent shown
in the gctrace output and makes this available programmatically.
This does make one small effect on the gctrace output: we now report
the duration of mark termination up to just before the final
start-the-world, rather than up to just after. However, unlike
stop-the-world, I don't believe there's any way that start-the-world
can block, so it should take negligible time.
While there are many statistics one might want to expose via MemStats,
this is one of the few that will undoubtedly remain meaningful
regardless of future changes to the memory system.
The diff for this change is larger than the actual change. Mostly it
lifts the code for computing the GC CPU utilization out of the
debug.gctrace path.
Updates #10323.
Change-Id: I0f7dc3fdcafe95e8d1233ceb79de606b48acd989
Reviewed-on: https://go-review.googlesource.com/12844
Reviewed-by: Russ Cox <rsc@golang.org>
Currently we only capture GC phase transition times if
debug.gctrace>0, but we're about to compute GC CPU utilization
regardless of whether debug.gctrace is set, so we need these
regardless of debug.gctrace.
Change-Id: If3acf16505a43d416e9a99753206f03287180660
Reviewed-on: https://go-review.googlesource.com/12843
Reviewed-by: Russ Cox <rsc@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
The scheduler, work buffer's dispose, and write barriers
can conspire to hide the a pointer from the GC's concurent
mark phase. If this pointer is the only path to a large
amount of marking the STW mark termination phase may take
a lot of time.
Consider the following:
1) dispose places a work buffer on the partial queue
2) the GC is busy so it does not immediately remove and
process the work buffer
3) the scheduler runs a mutator whose write barrier dequeues the
work buffer from the partial queue so the GC won't see it
This repeats until the GC reaches the mark termination
phase where the GC finally discovers the pointer along
with a lot of work to do.
This CL fixes the problem by having the mutator
dispose of the buffer to the full queue instead of
the partial queue. Since the write buffer never asks for full
buffers the conspiracy described above is not possible.
Updates #11694.
Change-Id: I2ce832f9657a7570f800e8ce4459cd9e304ef43b
Reviewed-on: https://go-review.googlesource.com/12840
Reviewed-by: Austin Clements <austin@google.com>
Currently we enter mark 2 by first flushing all existing gcWork caches
and then setting gcBlackenPromptly, which disables further gcWork
caching. However, if a worker or assist pulls a work buffer in to its
gcWork cache after that cache has been flushed but before caching is
disabled, that work may remain in that cache until mark termination.
If that work represents a heap bottleneck (e.g., a single pointer that
is the only way to reach a large amount of the heap), this can force
mark termination to do a large amount of work, resulting in a long
STW.
Fix this by reversing the order of these steps: first disable caching,
then flush all existing caches.
Rick Hudson <rlh> did the hard work of tracking this down. This CL
combined with CL 12672 and CL 12646 distills the critical parts of his
fix from CL 12539.
Fixes#11694.
Change-Id: Ib10d0a21e3f6170a80727d0286f9990df049fed2
Reviewed-on: https://go-review.googlesource.com/12688
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently the GC coordinator enables GC assists at the same time it
enables background mark workers, after the concurrent scan phase is
done. However, this means a rapidly allocating mutator has the entire
scan phase during which to allocate beyond the heap trigger and
potentially beyond the heap goal with no back-pressure from assists.
This prevents the feedback system that's supposed to keep the heap
size under the heap goal from doing its job.
Fix this by enabling mutator assists during the scan phase. This is
safe because the write barrier is already enabled and globally
acknowledged at this point.
There's still a very small window between when the heap size reaches
the heap trigger and when the GC coordinator is able to stop the world
during which the mutator can allocate unabated. This allows *very*
rapidly allocator mutators like TestTraceStress to still occasionally
exceed the heap goal by a small amount (~20 MB at most for
TestTraceStress). However, this seems like a corner case.
Fixes#11677.
Change-Id: I0f80d949ec82341cd31ca1604a626efb7295a819
Reviewed-on: https://go-review.googlesource.com/12674
Reviewed-by: Russ Cox <rsc@golang.org>
