Mon, 27 Jan 2014 13:14:53 +0100
8030177: G1: Enable TLAB resizing
Reviewed-by: tschatzl, stefank, jmasa
1 /*
2 * Copyright (c) 1999, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "memory/genCollectedHeap.hpp"
27 #include "memory/resourceArea.hpp"
28 #include "memory/threadLocalAllocBuffer.inline.hpp"
29 #include "memory/universe.inline.hpp"
30 #include "oops/oop.inline.hpp"
31 #include "runtime/thread.inline.hpp"
32 #include "utilities/copy.hpp"
34 // Thread-Local Edens support
36 // static member initialization
37 size_t ThreadLocalAllocBuffer::_max_size = 0;
38 unsigned ThreadLocalAllocBuffer::_target_refills = 0;
39 GlobalTLABStats* ThreadLocalAllocBuffer::_global_stats = NULL;
41 void ThreadLocalAllocBuffer::clear_before_allocation() {
42 _slow_refill_waste += (unsigned)remaining();
43 make_parsable(true); // also retire the TLAB
44 }
46 void ThreadLocalAllocBuffer::accumulate_statistics_before_gc() {
47 global_stats()->initialize();
49 for (JavaThread *thread = Threads::first(); thread != NULL; thread = thread->next()) {
50 thread->tlab().accumulate_statistics();
51 thread->tlab().initialize_statistics();
52 }
54 // Publish new stats if some allocation occurred.
55 if (global_stats()->allocation() != 0) {
56 global_stats()->publish();
57 if (PrintTLAB) {
58 global_stats()->print();
59 }
60 }
61 }
63 void ThreadLocalAllocBuffer::accumulate_statistics() {
64 Thread* thread = myThread();
65 size_t capacity = Universe::heap()->tlab_capacity(thread);
66 size_t used = Universe::heap()->tlab_used(thread);
68 _gc_waste += (unsigned)remaining();
69 size_t total_allocated = thread->allocated_bytes();
70 size_t allocated_since_last_gc = total_allocated - _allocated_before_last_gc;
71 _allocated_before_last_gc = total_allocated;
73 if (PrintTLAB && (_number_of_refills > 0 || Verbose)) {
74 print_stats("gc");
75 }
77 if (_number_of_refills > 0) {
78 // Update allocation history if a reasonable amount of eden was allocated.
79 bool update_allocation_history = used > 0.5 * capacity;
81 if (update_allocation_history) {
82 // Average the fraction of eden allocated in a tlab by this
83 // thread for use in the next resize operation.
84 // _gc_waste is not subtracted because it's included in
85 // "used".
86 // The result can be larger than 1.0 due to direct to old allocations.
87 // These allocations should ideally not be counted but since it is not possible
88 // to filter them out here we just cap the fraction to be at most 1.0.
89 double alloc_frac = MIN2(1.0, (double) allocated_since_last_gc / used);
90 _allocation_fraction.sample(alloc_frac);
91 }
92 global_stats()->update_allocating_threads();
93 global_stats()->update_number_of_refills(_number_of_refills);
94 global_stats()->update_allocation(_number_of_refills * desired_size());
95 global_stats()->update_gc_waste(_gc_waste);
96 global_stats()->update_slow_refill_waste(_slow_refill_waste);
97 global_stats()->update_fast_refill_waste(_fast_refill_waste);
99 } else {
100 assert(_number_of_refills == 0 && _fast_refill_waste == 0 &&
101 _slow_refill_waste == 0 && _gc_waste == 0,
102 "tlab stats == 0");
103 }
104 global_stats()->update_slow_allocations(_slow_allocations);
105 }
107 // Fills the current tlab with a dummy filler array to create
108 // an illusion of a contiguous Eden and optionally retires the tlab.
109 // Waste accounting should be done in caller as appropriate; see,
110 // for example, clear_before_allocation().
111 void ThreadLocalAllocBuffer::make_parsable(bool retire) {
112 if (end() != NULL) {
113 invariants();
115 if (retire) {
116 myThread()->incr_allocated_bytes(used_bytes());
117 }
119 CollectedHeap::fill_with_object(top(), hard_end(), retire);
121 if (retire || ZeroTLAB) { // "Reset" the TLAB
122 set_start(NULL);
123 set_top(NULL);
124 set_pf_top(NULL);
125 set_end(NULL);
126 }
127 }
128 assert(!(retire || ZeroTLAB) ||
129 (start() == NULL && end() == NULL && top() == NULL),
130 "TLAB must be reset");
131 }
133 void ThreadLocalAllocBuffer::resize_all_tlabs() {
134 if (ResizeTLAB) {
135 for (JavaThread *thread = Threads::first(); thread != NULL; thread = thread->next()) {
136 thread->tlab().resize();
137 }
138 }
139 }
141 void ThreadLocalAllocBuffer::resize() {
142 // Compute the next tlab size using expected allocation amount
143 assert(ResizeTLAB, "Should not call this otherwise");
144 size_t alloc = (size_t)(_allocation_fraction.average() *
145 (Universe::heap()->tlab_capacity(myThread()) / HeapWordSize));
146 size_t new_size = alloc / _target_refills;
148 new_size = MIN2(MAX2(new_size, min_size()), max_size());
150 size_t aligned_new_size = align_object_size(new_size);
152 if (PrintTLAB && Verbose) {
153 gclog_or_tty->print("TLAB new size: thread: " INTPTR_FORMAT " [id: %2d]"
154 " refills %d alloc: %8.6f desired_size: " SIZE_FORMAT " -> " SIZE_FORMAT "\n",
155 myThread(), myThread()->osthread()->thread_id(),
156 _target_refills, _allocation_fraction.average(), desired_size(), aligned_new_size);
157 }
158 set_desired_size(aligned_new_size);
159 set_refill_waste_limit(initial_refill_waste_limit());
160 }
162 void ThreadLocalAllocBuffer::initialize_statistics() {
163 _number_of_refills = 0;
164 _fast_refill_waste = 0;
165 _slow_refill_waste = 0;
166 _gc_waste = 0;
167 _slow_allocations = 0;
168 }
170 void ThreadLocalAllocBuffer::fill(HeapWord* start,
171 HeapWord* top,
172 size_t new_size) {
173 _number_of_refills++;
174 if (PrintTLAB && Verbose) {
175 print_stats("fill");
176 }
177 assert(top <= start + new_size - alignment_reserve(), "size too small");
178 initialize(start, top, start + new_size - alignment_reserve());
180 // Reset amount of internal fragmentation
181 set_refill_waste_limit(initial_refill_waste_limit());
182 }
184 void ThreadLocalAllocBuffer::initialize(HeapWord* start,
185 HeapWord* top,
186 HeapWord* end) {
187 set_start(start);
188 set_top(top);
189 set_pf_top(top);
190 set_end(end);
191 invariants();
192 }
194 void ThreadLocalAllocBuffer::initialize() {
195 initialize(NULL, // start
196 NULL, // top
197 NULL); // end
199 set_desired_size(initial_desired_size());
201 // Following check is needed because at startup the main (primordial)
202 // thread is initialized before the heap is. The initialization for
203 // this thread is redone in startup_initialization below.
204 if (Universe::heap() != NULL) {
205 size_t capacity = Universe::heap()->tlab_capacity(myThread()) / HeapWordSize;
206 double alloc_frac = desired_size() * target_refills() / (double) capacity;
207 _allocation_fraction.sample(alloc_frac);
208 }
210 set_refill_waste_limit(initial_refill_waste_limit());
212 initialize_statistics();
213 }
215 void ThreadLocalAllocBuffer::startup_initialization() {
217 // Assuming each thread's active tlab is, on average,
218 // 1/2 full at a GC
219 _target_refills = 100 / (2 * TLABWasteTargetPercent);
220 _target_refills = MAX2(_target_refills, (unsigned)1U);
222 _global_stats = new GlobalTLABStats();
224 // During jvm startup, the main (primordial) thread is initialized
225 // before the heap is initialized. So reinitialize it now.
226 guarantee(Thread::current()->is_Java_thread(), "tlab initialization thread not Java thread");
227 Thread::current()->tlab().initialize();
229 if (PrintTLAB && Verbose) {
230 gclog_or_tty->print("TLAB min: " SIZE_FORMAT " initial: " SIZE_FORMAT " max: " SIZE_FORMAT "\n",
231 min_size(), Thread::current()->tlab().initial_desired_size(), max_size());
232 }
233 }
235 size_t ThreadLocalAllocBuffer::initial_desired_size() {
236 size_t init_sz;
238 if (TLABSize > 0) {
239 init_sz = MIN2(TLABSize / HeapWordSize, max_size());
240 } else if (global_stats() == NULL) {
241 // Startup issue - main thread initialized before heap initialized.
242 init_sz = min_size();
243 } else {
244 // Initial size is a function of the average number of allocating threads.
245 unsigned nof_threads = global_stats()->allocating_threads_avg();
247 init_sz = (Universe::heap()->tlab_capacity(myThread()) / HeapWordSize) /
248 (nof_threads * target_refills());
249 init_sz = align_object_size(init_sz);
250 init_sz = MIN2(MAX2(init_sz, min_size()), max_size());
251 }
252 return init_sz;
253 }
255 void ThreadLocalAllocBuffer::print_stats(const char* tag) {
256 Thread* thrd = myThread();
257 size_t waste = _gc_waste + _slow_refill_waste + _fast_refill_waste;
258 size_t alloc = _number_of_refills * _desired_size;
259 double waste_percent = alloc == 0 ? 0.0 :
260 100.0 * waste / alloc;
261 size_t tlab_used = Universe::heap()->tlab_used(thrd);
262 gclog_or_tty->print("TLAB: %s thread: " INTPTR_FORMAT " [id: %2d]"
263 " desired_size: " SIZE_FORMAT "KB"
264 " slow allocs: %d refill waste: " SIZE_FORMAT "B"
265 " alloc:%8.5f %8.0fKB refills: %d waste %4.1f%% gc: %dB"
266 " slow: %dB fast: %dB\n",
267 tag, thrd, thrd->osthread()->thread_id(),
268 _desired_size / (K / HeapWordSize),
269 _slow_allocations, _refill_waste_limit * HeapWordSize,
270 _allocation_fraction.average(),
271 _allocation_fraction.average() * tlab_used / K,
272 _number_of_refills, waste_percent,
273 _gc_waste * HeapWordSize,
274 _slow_refill_waste * HeapWordSize,
275 _fast_refill_waste * HeapWordSize);
276 }
278 void ThreadLocalAllocBuffer::verify() {
279 HeapWord* p = start();
280 HeapWord* t = top();
281 HeapWord* prev_p = NULL;
282 while (p < t) {
283 oop(p)->verify();
284 prev_p = p;
285 p += oop(p)->size();
286 }
287 guarantee(p == top(), "end of last object must match end of space");
288 }
290 Thread* ThreadLocalAllocBuffer::myThread() {
291 return (Thread*)(((char *)this) +
292 in_bytes(start_offset()) -
293 in_bytes(Thread::tlab_start_offset()));
294 }
297 GlobalTLABStats::GlobalTLABStats() :
298 _allocating_threads_avg(TLABAllocationWeight) {
300 initialize();
302 _allocating_threads_avg.sample(1); // One allocating thread at startup
304 if (UsePerfData) {
306 EXCEPTION_MARK;
307 ResourceMark rm;
309 char* cname = PerfDataManager::counter_name("tlab", "allocThreads");
310 _perf_allocating_threads =
311 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
313 cname = PerfDataManager::counter_name("tlab", "fills");
314 _perf_total_refills =
315 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
317 cname = PerfDataManager::counter_name("tlab", "maxFills");
318 _perf_max_refills =
319 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
321 cname = PerfDataManager::counter_name("tlab", "alloc");
322 _perf_allocation =
323 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
325 cname = PerfDataManager::counter_name("tlab", "gcWaste");
326 _perf_gc_waste =
327 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
329 cname = PerfDataManager::counter_name("tlab", "maxGcWaste");
330 _perf_max_gc_waste =
331 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
333 cname = PerfDataManager::counter_name("tlab", "slowWaste");
334 _perf_slow_refill_waste =
335 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
337 cname = PerfDataManager::counter_name("tlab", "maxSlowWaste");
338 _perf_max_slow_refill_waste =
339 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
341 cname = PerfDataManager::counter_name("tlab", "fastWaste");
342 _perf_fast_refill_waste =
343 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
345 cname = PerfDataManager::counter_name("tlab", "maxFastWaste");
346 _perf_max_fast_refill_waste =
347 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
349 cname = PerfDataManager::counter_name("tlab", "slowAlloc");
350 _perf_slow_allocations =
351 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
353 cname = PerfDataManager::counter_name("tlab", "maxSlowAlloc");
354 _perf_max_slow_allocations =
355 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
356 }
357 }
359 void GlobalTLABStats::initialize() {
360 // Clear counters summarizing info from all threads
361 _allocating_threads = 0;
362 _total_refills = 0;
363 _max_refills = 0;
364 _total_allocation = 0;
365 _total_gc_waste = 0;
366 _max_gc_waste = 0;
367 _total_slow_refill_waste = 0;
368 _max_slow_refill_waste = 0;
369 _total_fast_refill_waste = 0;
370 _max_fast_refill_waste = 0;
371 _total_slow_allocations = 0;
372 _max_slow_allocations = 0;
373 }
375 void GlobalTLABStats::publish() {
376 _allocating_threads_avg.sample(_allocating_threads);
377 if (UsePerfData) {
378 _perf_allocating_threads ->set_value(_allocating_threads);
379 _perf_total_refills ->set_value(_total_refills);
380 _perf_max_refills ->set_value(_max_refills);
381 _perf_allocation ->set_value(_total_allocation);
382 _perf_gc_waste ->set_value(_total_gc_waste);
383 _perf_max_gc_waste ->set_value(_max_gc_waste);
384 _perf_slow_refill_waste ->set_value(_total_slow_refill_waste);
385 _perf_max_slow_refill_waste->set_value(_max_slow_refill_waste);
386 _perf_fast_refill_waste ->set_value(_total_fast_refill_waste);
387 _perf_max_fast_refill_waste->set_value(_max_fast_refill_waste);
388 _perf_slow_allocations ->set_value(_total_slow_allocations);
389 _perf_max_slow_allocations ->set_value(_max_slow_allocations);
390 }
391 }
393 void GlobalTLABStats::print() {
394 size_t waste = _total_gc_waste + _total_slow_refill_waste + _total_fast_refill_waste;
395 double waste_percent = _total_allocation == 0 ? 0.0 :
396 100.0 * waste / _total_allocation;
397 gclog_or_tty->print("TLAB totals: thrds: %d refills: %d max: %d"
398 " slow allocs: %d max %d waste: %4.1f%%"
399 " gc: " SIZE_FORMAT "B max: " SIZE_FORMAT "B"
400 " slow: " SIZE_FORMAT "B max: " SIZE_FORMAT "B"
401 " fast: " SIZE_FORMAT "B max: " SIZE_FORMAT "B\n",
402 _allocating_threads,
403 _total_refills, _max_refills,
404 _total_slow_allocations, _max_slow_allocations,
405 waste_percent,
406 _total_gc_waste * HeapWordSize,
407 _max_gc_waste * HeapWordSize,
408 _total_slow_refill_waste * HeapWordSize,
409 _max_slow_refill_waste * HeapWordSize,
410 _total_fast_refill_waste * HeapWordSize,
411 _max_fast_refill_waste * HeapWordSize);
412 }