< prev index next >
src/share/vm/gc_implementation/shenandoah/shenandoahPacer.cpp
Print this page
rev 10511 : Fix MacOS/Clang build failure
rev 10543 : [backport] Application pacing precision fixes
rev 10558 : [backport] Exponential backoff with pacing
rev 10559 : [backport] More detailed pacing histogram
rev 10563 : [backport] Pacer for evacuation should print "Avail" to capture discounting
rev 10566 : [backport] Pacer should account actual size for elastic TLABs
rev 10575 : Fix x86_32 build
rev 10599 : [backport] Convert magic value to ShenandoahPacingSurcharge
rev 10620 : [backport] Evac reserve: make sure GC has untouchable space to move the objects into
@@ -33,16 +33,22 @@
*
* Here, we do not know how large would be the collection set, and what are the
* relative performances of the each stage in the concurrent cycle, and so we have to
* make some assumptions.
*
- * We assume, for pessimistic reasons, that the entire heap is full of alive objects,
- * and it will be evacuated fully. Therefore, we count live objects visited by all three
- * stages against the heap used at the beginning of the collection. That means if there
- * are dead objects, they would not be accounted for in this budget, and that would mean
- * allocation would be pacified excessively. But that *also* means the collection cycle
- * would finish earlier than pacer expects.
+ * For concurrent mark, there is no clear notion of progress. The moderately accurate
+ * and easy to get metric is the amount of live objects the mark had encountered. But,
+ * that does directly correlate with the used heap, because the heap might be fully
+ * dead or fully alive. We cannot assume either of the extremes: we would either allow
+ * application to run out of memory if we assume heap is fully dead but it is not, and,
+ * conversely, we would pacify application excessively if we assume heap is fully alive
+ * but it is not. So we need to guesstimate the particular expected value for heap liveness.
+ * The best way to do this is apparently recording the past history.
+ *
+ * For concurrent evac and update-refs, we are walking the heap per-region, and so the
+ * notion of progress is clear: we get reported the "used" size from the processed regions
+ * and use the global heap-used as the baseline.
*
* The allocatable space when GC is running is "free" at the start of cycle, but the
* accounted budget is based on "used". So, we need to adjust the tax knowing that.
* Also, since we effectively count the used space three times (mark, evac, update-refs),
* we need to multiply the tax by 3. Example: for 10 MB free and 90 MB used, GC would
@@ -51,47 +57,46 @@
*/
void ShenandoahPacer::setup_for_mark() {
assert(ShenandoahPacing, "Only be here when pacing is enabled");
- size_t used = _heap->used();
+ size_t live = update_and_get_progress_history();
size_t free = _heap->free_set()->available();
size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
size_t taxable = free - non_taxable;
- double tax = 1.0 * used / taxable; // base tax for available free space
+ double tax = 1.0 * live / taxable; // base tax for available free space
tax *= 3; // mark is phase 1 of 3, claim 1/3 of free for it
- tax = MAX2<double>(1, tax); // never allocate more than GC collects during the cycle
- tax *= 1.1; // additional surcharge to help unclutter heap
+ tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap
restart_with(non_taxable, tax);
- log_info(gc, ergo)("Pacer for Mark. Used: " SIZE_FORMAT "M, Free: " SIZE_FORMAT
+ log_info(gc, ergo)("Pacer for Mark. Expected Live: " SIZE_FORMAT "M, Free: " SIZE_FORMAT
"M, Non-Taxable: " SIZE_FORMAT "M, Alloc Tax Rate: %.1fx",
- used / M, free / M, non_taxable / M, tax);
+ live / M, free / M, non_taxable / M, tax);
}
void ShenandoahPacer::setup_for_evac() {
assert(ShenandoahPacing, "Only be here when pacing is enabled");
- size_t cset = _heap->collection_set()->live_data();
+ size_t used = _heap->collection_set()->used();
size_t free = _heap->free_set()->available();
size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
size_t taxable = free - non_taxable;
- double tax = 1.0 * cset / taxable; // base tax for available free space
+ double tax = 1.0 * used / taxable; // base tax for available free space
tax *= 2; // evac is phase 2 of 3, claim 1/2 of remaining free
- tax = MAX2<double>(1, tax); // never allocate more than GC collects during the cycle
- tax *= 1.1; // additional surcharge to help unclutter heap
+ tax = MAX2<double>(1, tax); // never allocate more than GC processes during the phase
+ tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap
restart_with(non_taxable, tax);
- log_info(gc, ergo)("Pacer for Evacuation. CSet: " SIZE_FORMAT "M, Free: " SIZE_FORMAT
+ log_info(gc, ergo)("Pacer for Evacuation. Used CSet: " SIZE_FORMAT "M, Free: " SIZE_FORMAT
"M, Non-Taxable: " SIZE_FORMAT "M, Alloc Tax Rate: %.1fx",
- cset / M, free / M, non_taxable / M, tax);
+ used / M, free / M, non_taxable / M, tax);
}
void ShenandoahPacer::setup_for_updaterefs() {
assert(ShenandoahPacing, "Only be here when pacing is enabled");
@@ -101,16 +106,16 @@
size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
size_t taxable = free - non_taxable;
double tax = 1.0 * used / taxable; // base tax for available free space
tax *= 1; // update-refs is phase 3 of 3, claim the remaining free
- tax = MAX2<double>(1, tax); // never allocate more than GC collects during the cycle
- tax *= 1.1; // additional surcharge to help unclutter heap
+ tax = MAX2<double>(1, tax); // never allocate more than GC processes during the phase
+ tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap
restart_with(non_taxable, tax);
- log_info(gc, ergo)("Pacer for Update-Refs. Used: " SIZE_FORMAT "M, Free: " SIZE_FORMAT
+ log_info(gc, ergo)("Pacer for Update Refs. Used: " SIZE_FORMAT "M, Free: " SIZE_FORMAT
"M, Non-Taxable: " SIZE_FORMAT "M, Alloc Tax Rate: %.1fx",
used / M, free / M, non_taxable / M, tax);
}
/*
@@ -132,24 +137,48 @@
log_info(gc, ergo)("Pacer for Idle. Initial: " SIZE_FORMAT "M, Alloc Tax Rate: %.1fx",
initial / M, tax);
}
+size_t ShenandoahPacer::update_and_get_progress_history() {
+ if (_progress == -1) {
+ // First initialization, report some prior
+ Atomic::store((intptr_t)PACING_PROGRESS_ZERO, &_progress);
+ return (size_t) (_heap->capacity() * 0.1);
+ } else {
+ // Record history, and reply historical data
+ _progress_history->add(_progress);
+ Atomic::store((intptr_t)PACING_PROGRESS_ZERO, &_progress);
+ return (size_t) (_progress_history->avg() * HeapWordSize);
+ }
+}
+
void ShenandoahPacer::restart_with(jlong non_taxable_bytes, jdouble tax_rate) {
STATIC_ASSERT(sizeof(size_t) <= sizeof(intptr_t));
- intptr_t initial = (size_t)(non_taxable_bytes * tax_rate) >> LogHeapWordSize;
+ {
+ intptr_t initial = (size_t) (non_taxable_bytes * tax_rate) >> LogHeapWordSize;
intptr_t cur;
do {
cur = OrderAccess::load_acquire(&_budget);
} while (Atomic::cmpxchg(initial, &_budget, cur) != cur);
+ }
+
OrderAccess::release_store(&_tax_rate, tax_rate);
+
+ {
+ intptr_t cur, val;
+ do {
+ cur = OrderAccess::load_acquire(&_epoch);
+ val = cur + 1;
+ } while (Atomic::cmpxchg(val, &_epoch, cur) != cur);
+ }
}
bool ShenandoahPacer::claim_for_alloc(size_t words, bool force) {
assert(ShenandoahPacing, "Only be here when pacing is enabled");
- intptr_t tax = MAX2<intptr_t>(1, words * OrderAccess::load_acquire(&_tax_rate));
+ intptr_t tax = MAX2<intptr_t>(1, (intptr_t)(words * OrderAccess::load_acquire(&_tax_rate)));
intptr_t cur = 0;
intptr_t new_val = 0;
do {
cur = OrderAccess::load_acquire(&_budget);
@@ -160,43 +189,71 @@
new_val = cur - tax;
} while (Atomic::cmpxchg(new_val, &_budget, cur) != cur);
return true;
}
+void ShenandoahPacer::unpace_for_alloc(intptr_t epoch, size_t words) {
+ assert(ShenandoahPacing, "Only be here when pacing is enabled");
+
+ if (_epoch != epoch) {
+ // Stale ticket, no need to unpace.
+ return;
+ }
+
+ intptr_t tax = MAX2<intptr_t>(1, (intptr_t)(words * OrderAccess::load_acquire(&_tax_rate)));
+ Atomic::add(tax, &_budget);
+}
+
+intptr_t ShenandoahPacer::epoch() {
+ return OrderAccess::load_acquire(&_epoch);
+}
+
void ShenandoahPacer::pace_for_alloc(size_t words) {
assert(ShenandoahPacing, "Only be here when pacing is enabled");
// Fast path: try to allocate right away
if (claim_for_alloc(words, false)) {
return;
}
- size_t max_wait_ms = ShenandoahPacingMaxDelay;
+ size_t max = ShenandoahPacingMaxDelay;
double start = os::elapsedTime();
+ size_t total = 0;
+ size_t cur = 0;
+
while (true) {
// We could instead assist GC, but this would suffice for now.
// This code should also participate in safepointing.
- os::sleep(Thread::current(), 1, true);
+ // Perform the exponential backoff, limited by max.
+
+ cur = cur * 2;
+ if (total + cur > max) {
+ cur = (max > total) ? (max - total) : 0;
+ }
+ cur = MAX2<size_t>(1, cur);
+
+ os::sleep(Thread::current(), cur, true);
double end = os::elapsedTime();
- size_t ms = (size_t)((end - start) * 1000);
- if (ms > max_wait_ms) {
+ total = (size_t)((end - start) * 1000);
+
+ if (total > max) {
// Spent local time budget to wait for enough GC progress.
// Breaking out and allocating anyway, which may mean we outpace GC,
// and start Degenerated GC cycle.
- _delays.add(ms);
+ _delays.add(total);
// Forcefully claim the budget: it may go negative at this point, and
// GC should replenish for this and subsequent allocations
claim_for_alloc(words, true);
break;
}
if (claim_for_alloc(words, false)) {
// Acquired enough permit, nice. Can allocate now.
- _delays.add(ms);
+ _delays.add(total);
break;
}
}
}
@@ -214,12 +271,22 @@
out->cr();
out->print_cr("Actual pacing delays histogram:");
out->cr();
- out->print_cr("%10s - %10s %12s", "From", "To", "Count");
+ out->print_cr("%10s - %10s %12s%12s", "From", "To", "Count", "Sum");
+
+ size_t total_count = 0;
+ size_t total_sum = 0;
for (int c = _delays.min_level(); c <= _delays.max_level(); c++) {
- out->print("%7d ms - %7d ms:", (c == 0) ? 0 : 1 << (c - 1), 1 << c);
- out->print_cr(SIZE_FORMAT_W(12), _delays.level(c));
+ int l = (c == 0) ? 0 : 1 << (c - 1);
+ int r = 1 << c;
+ size_t count = _delays.level(c);
+ size_t sum = count * (r - l) / 2;
+ total_count += count;
+ total_sum += sum;
+
+ out->print_cr("%7d ms - %7d ms: " SIZE_FORMAT_W(12) SIZE_FORMAT_W(12) " ms", l, r, count, sum);
}
+ out->print_cr("%23s: " SIZE_FORMAT_W(12) SIZE_FORMAT_W(12) " ms", "Total", total_count, total_sum);
out->cr();
}
< prev index next >