1 /* 2 * Copyright (c) 2010, 2016, 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 */ 24 25 #ifndef SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP 26 #define SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP 27 28 #include "runtime/simpleThresholdPolicy.hpp" 29 30 #ifdef TIERED 31 class CompileTask; 32 class CompileQueue; 33 34 /* 35 * The system supports 5 execution levels: 36 * * level 0 - interpreter 37 * * level 1 - C1 with full optimization (no profiling) 38 * * level 2 - C1 with invocation and backedge counters 39 * * level 3 - C1 with full profiling (level 2 + MDO) 40 * * level 4 - C2 41 * 42 * Levels 0, 2 and 3 periodically notify the runtime about the current value of the counters 43 * (invocation counters and backedge counters). The frequency of these notifications is 44 * different at each level. These notifications are used by the policy to decide what transition 45 * to make. 46 * 47 * Execution starts at level 0 (interpreter), then the policy can decide either to compile the 48 * method at level 3 or level 2. The decision is based on the following factors: 49 * 1. The length of the C2 queue determines the next level. The observation is that level 2 50 * is generally faster than level 3 by about 30%, therefore we would want to minimize the time 51 * a method spends at level 3. We should only spend the time at level 3 that is necessary to get 52 * adequate profiling. So, if the C2 queue is long enough it is more beneficial to go first to 53 * level 2, because if we transitioned to level 3 we would be stuck there until our C2 compile 54 * request makes its way through the long queue. When the load on C2 recedes we are going to 55 * recompile at level 3 and start gathering profiling information. 56 * 2. The length of C1 queue is used to dynamically adjust the thresholds, so as to introduce 57 * additional filtering if the compiler is overloaded. The rationale is that by the time a 58 * method gets compiled it can become unused, so it doesn't make sense to put too much onto the 59 * queue. 60 * 61 * After profiling is completed at level 3 the transition is made to level 4. Again, the length 62 * of the C2 queue is used as a feedback to adjust the thresholds. 63 * 64 * After the first C1 compile some basic information is determined about the code like the number 65 * of the blocks and the number of the loops. Based on that it can be decided that a method 66 * is trivial and compiling it with C1 will yield the same code. In this case the method is 67 * compiled at level 1 instead of 4. 68 * 69 * We also support profiling at level 0. If C1 is slow enough to produce the level 3 version of 70 * the code and the C2 queue is sufficiently small we can decide to start profiling in the 71 * interpreter (and continue profiling in the compiled code once the level 3 version arrives). 72 * If the profiling at level 0 is fully completed before level 3 version is produced, a level 2 73 * version is compiled instead in order to run faster waiting for a level 4 version. 74 * 75 * Compile queues are implemented as priority queues - for each method in the queue we compute 76 * the event rate (the number of invocation and backedge counter increments per unit of time). 77 * When getting an element off the queue we pick the one with the largest rate. Maintaining the 78 * rate also allows us to remove stale methods (the ones that got on the queue but stopped 79 * being used shortly after that). 80 */ 81 82 /* Command line options: 83 * - Tier?InvokeNotifyFreqLog and Tier?BackedgeNotifyFreqLog control the frequency of method 84 * invocation and backedge notifications. Basically every n-th invocation or backedge a mutator thread 85 * makes a call into the runtime. 86 * 87 * - Tier?InvocationThreshold, Tier?CompileThreshold, Tier?BackEdgeThreshold, Tier?MinInvocationThreshold control 88 * compilation thresholds. 89 * Level 2 thresholds are not used and are provided for option-compatibility and potential future use. 90 * Other thresholds work as follows: 91 * 92 * Transition from interpreter (level 0) to C1 with full profiling (level 3) happens when 93 * the following predicate is true (X is the level): 94 * 95 * i > TierXInvocationThreshold * s || (i > TierXMinInvocationThreshold * s && i + b > TierXCompileThreshold * s), 96 * 97 * where $i$ is the number of method invocations, $b$ number of backedges and $s$ is the scaling 98 * coefficient that will be discussed further. 99 * The intuition is to equalize the time that is spend profiling each method. 100 * The same predicate is used to control the transition from level 3 to level 4 (C2). It should be 101 * noted though that the thresholds are relative. Moreover i and b for the 0->3 transition come 102 * from Method* and for 3->4 transition they come from MDO (since profiled invocations are 103 * counted separately). Finally, if a method does not contain anything worth profiling, a transition 104 * from level 3 to level 4 occurs without considering thresholds (e.g., with fewer invocations than 105 * what is specified by Tier4InvocationThreshold). 106 * 107 * OSR transitions are controlled simply with b > TierXBackEdgeThreshold * s predicates. 108 * 109 * - Tier?LoadFeedback options are used to automatically scale the predicates described above depending 110 * on the compiler load. The scaling coefficients are computed as follows: 111 * 112 * s = queue_size_X / (TierXLoadFeedback * compiler_count_X) + 1, 113 * 114 * where queue_size_X is the current size of the compiler queue of level X, and compiler_count_X 115 * is the number of level X compiler threads. 116 * 117 * Basically these parameters describe how many methods should be in the compile queue 118 * per compiler thread before the scaling coefficient increases by one. 119 * 120 * This feedback provides the mechanism to automatically control the flow of compilation requests 121 * depending on the machine speed, mutator load and other external factors. 122 * 123 * - Tier3DelayOn and Tier3DelayOff parameters control another important feedback loop. 124 * Consider the following observation: a method compiled with full profiling (level 3) 125 * is about 30% slower than a method at level 2 (just invocation and backedge counters, no MDO). 126 * Normally, the following transitions will occur: 0->3->4. The problem arises when the C2 queue 127 * gets congested and the 3->4 transition is delayed. While the method is the C2 queue it continues 128 * executing at level 3 for much longer time than is required by the predicate and at suboptimal speed. 129 * The idea is to dynamically change the behavior of the system in such a way that if a substantial 130 * load on C2 is detected we would first do the 0->2 transition allowing a method to run faster. 131 * And then when the load decreases to allow 2->3 transitions. 132 * 133 * Tier3Delay* parameters control this switching mechanism. 134 * Tier3DelayOn is the number of methods in the C2 queue per compiler thread after which the policy 135 * no longer does 0->3 transitions but does 0->2 transitions instead. 136 * Tier3DelayOff switches the original behavior back when the number of methods in the C2 queue 137 * per compiler thread falls below the specified amount. 138 * The hysteresis is necessary to avoid jitter. 139 * 140 * - TieredCompileTaskTimeout is the amount of time an idle method can spend in the compile queue. 141 * Basically, since we use the event rate d(i + b)/dt as a value of priority when selecting a method to 142 * compile from the compile queue, we also can detect stale methods for which the rate has been 143 * 0 for some time in the same iteration. Stale methods can appear in the queue when an application 144 * abruptly changes its behavior. 145 * 146 * - TieredStopAtLevel, is used mostly for testing. It allows to bypass the policy logic and stick 147 * to a given level. For example it's useful to set TieredStopAtLevel = 1 in order to compile everything 148 * with pure c1. 149 * 150 * - Tier0ProfilingStartPercentage allows the interpreter to start profiling when the inequalities in the 151 * 0->3 predicate are already exceeded by the given percentage but the level 3 version of the 152 * method is still not ready. We can even go directly from level 0 to 4 if c1 doesn't produce a compiled 153 * version in time. This reduces the overall transition to level 4 and decreases the startup time. 154 * Note that this behavior is also guarded by the Tier3Delay mechanism: when the c2 queue is too long 155 * these is not reason to start profiling prematurely. 156 * 157 * - TieredRateUpdateMinTime and TieredRateUpdateMaxTime are parameters of the rate computation. 158 * Basically, the rate is not computed more frequently than TieredRateUpdateMinTime and is considered 159 * to be zero if no events occurred in TieredRateUpdateMaxTime. 160 */ 161 162 163 class AdvancedThresholdPolicy : public SimpleThresholdPolicy { 164 jlong _start_time; 165 166 // Call and loop predicates determine whether a transition to a higher compilation 167 // level should be performed (pointers to predicate functions are passed to common(). 168 // Predicates also take compiler load into account. 169 typedef bool (AdvancedThresholdPolicy::*Predicate)(int i, int b, CompLevel cur_level, Method* method); 170 bool call_predicate(int i, int b, CompLevel cur_level, Method* method); 171 bool loop_predicate(int i, int b, CompLevel cur_level, Method* method); 172 // Common transition function. Given a predicate determines if a method should transition to another level. 173 CompLevel common(Predicate p, Method* method, CompLevel cur_level, bool disable_feedback = false); 174 // Transition functions. 175 // call_event determines if a method should be compiled at a different 176 // level with a regular invocation entry. 177 CompLevel call_event(Method* method, CompLevel cur_level, JavaThread * thread); 178 // loop_event checks if a method should be OSR compiled at a different 179 // level. 180 CompLevel loop_event(Method* method, CompLevel cur_level, JavaThread * thread); 181 // Has a method been long around? 182 // We don't remove old methods from the compile queue even if they have 183 // very low activity (see select_task()). 184 inline bool is_old(Method* method); 185 // Was a given method inactive for a given number of milliseconds. 186 // If it is, we would remove it from the queue (see select_task()). 187 inline bool is_stale(jlong t, jlong timeout, Method* m); 188 // Compute the weight of the method for the compilation scheduling 189 inline double weight(Method* method); 190 // Apply heuristics and return true if x should be compiled before y 191 inline bool compare_methods(Method* x, Method* y); 192 // Compute event rate for a given method. The rate is the number of event (invocations + backedges) 193 // per millisecond. 194 inline void update_rate(jlong t, Method* m); 195 // Compute threshold scaling coefficient 196 inline double threshold_scale(CompLevel level, int feedback_k); 197 // If a method is old enough and is still in the interpreter we would want to 198 // start profiling without waiting for the compiled method to arrive. This function 199 // determines whether we should do that. 200 inline bool should_create_mdo(Method* method, CompLevel cur_level); 201 // Create MDO if necessary. 202 void create_mdo(methodHandle mh, JavaThread* thread); 203 // Is method profiled enough? 204 bool is_method_profiled(Method* method); 205 206 double _increase_threshold_at_ratio; 207 208 bool maybe_switch_to_aot(methodHandle mh, CompLevel cur_level, CompLevel next_level, JavaThread* thread); 209 210 protected: 211 void print_specific(EventType type, methodHandle mh, methodHandle imh, int bci, CompLevel level); 212 213 void set_increase_threshold_at_ratio() { _increase_threshold_at_ratio = 100 / (100 - (double)IncreaseFirstTierCompileThresholdAt); } 214 void set_start_time(jlong t) { _start_time = t; } 215 jlong start_time() const { return _start_time; } 216 217 // Submit a given method for compilation (and update the rate). 218 virtual void submit_compile(const methodHandle& mh, int bci, CompLevel level, JavaThread* thread); 219 // event() from SimpleThresholdPolicy would call these. 220 virtual void method_invocation_event(const methodHandle& method, const methodHandle& inlinee, 221 CompLevel level, CompiledMethod* nm, JavaThread* thread); 222 virtual void method_back_branch_event(const methodHandle& method, const methodHandle& inlinee, 223 int bci, CompLevel level, CompiledMethod* nm, JavaThread* thread); 224 public: 225 AdvancedThresholdPolicy() : _start_time(0) { } 226 // Select task is called by CompileBroker. We should return a task or NULL. 227 virtual CompileTask* select_task(CompileQueue* compile_queue); 228 virtual void initialize(); 229 virtual bool should_not_inline(ciEnv* env, ciMethod* callee); 230 231 }; 232 233 #endif // TIERED 234 235 #endif // SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP