1 /* 2 * Copyright (c) 2010, 2011 Oracle and/or its affiliates. All rights reserved. 3 * ORACLE PROPRIETARY/CONFIDENTIAL. Use is subject to license terms. 4 */ 5 6 #ifndef SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP 7 #define SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP 8 9 #include "runtime/simpleThresholdPolicy.hpp" 10 11 #ifdef TIERED 12 class CompileTask; 13 class CompileQueue; 14 15 /* 16 * The system supports 5 execution levels: 17 * * level 0 - interpreter 18 * * level 1 - C1 with full optimization (no profiling) 19 * * level 2 - C1 with invocation and backedge counters 20 * * level 3 - C1 with full profiling (level 2 + MDO) 21 * * level 4 - C2 22 * 23 * Levels 0, 2 and 3 periodically notify the runtime about the current value of the counters 24 * (invocation counters and backedge counters). The frequency of these notifications is 25 * different at each level. These notifications are used by the policy to decide what transition 26 * to make. 27 * 28 * Execution starts at level 0 (interpreter), then the policy can decide either to compile the 29 * method at level 3 or level 2. The decision is based on the following factors: 30 * 1. The length of the C2 queue determines the next level. The observation is that level 2 31 * is generally faster than level 3 by about 30%, therefore we would want to minimize the time 32 * a method spends at level 3. We should only spend the time at level 3 that is necessary to get 33 * adequate profiling. So, if the C2 queue is long enough it is more beneficial to go first to 34 * level 2, because if we transitioned to level 3 we would be stuck there until our C2 compile 35 * request makes its way through the long queue. When the load on C2 recedes we are going to 36 * recompile at level 3 and start gathering profiling information. 37 * 2. The length of C1 queue is used to dynamically adjust the thresholds, so as to introduce 38 * additional filtering if the compiler is overloaded. The rationale is that by the time a 39 * method gets compiled it can become unused, so it doesn't make sense to put too much onto the 40 * queue. 41 * 42 * After profiling is completed at level 3 the transition is made to level 4. Again, the length 43 * of the C2 queue is used as a feedback to adjust the thresholds. 44 * 45 * After the first C1 compile some basic information is determined about the code like the number 46 * of the blocks and the number of the loops. Based on that it can be decided that a method 47 * is trivial and compiling it with C1 will yield the same code. In this case the method is 48 * compiled at level 1 instead of 4. 49 * 50 * We also support profiling at level 0. If C1 is slow enough to produce the level 3 version of 51 * the code and the C2 queue is sufficiently small we can decide to start profiling in the 52 * interpreter (and continue profiling in the compiled code once the level 3 version arrives). 53 * If the profiling at level 0 is fully completed before level 3 version is produced, a level 2 54 * version is compiled instead in order to run faster waiting for a level 4 version. 55 * 56 * Compile queues are implemented as priority queues - for each method in the queue we compute 57 * the event rate (the number of invocation and backedge counter increments per unit of time). 58 * When getting an element off the queue we pick the one with the largest rate. Maintaining the 59 * rate also allows us to remove stale methods (the ones that got on the queue but stopped 60 * being used shortly after that). 61 */ 62 63 /* Command line options: 64 * - Tier?InvokeNotifyFreqLog and Tier?BackedgeNotifyFreqLog control the frequency of method 65 * invocation and backedge notifications. Basically every n-th invocation or backedge a mutator thread 66 * makes a call into the runtime. 67 * 68 * - Tier?CompileThreshold, Tier?BackEdgeThreshold, Tier?MinInvocationThreshold control 69 * compilation thresholds. 70 * Level 2 thresholds are not used and are provided for option-compatibility and potential future use. 71 * Other thresholds work as follows: 72 * 73 * Transition from interpreter (level 0) to C1 with full profiling (level 3) happens when 74 * the following predicate is true (X is the level): 75 * 76 * i > TierXInvocationThreshold * s || (i > TierXMinInvocationThreshold * s && i + b > TierXCompileThreshold * s), 77 * 78 * where $i$ is the number of method invocations, $b$ number of backedges and $s$ is the scaling 79 * coefficient that will be discussed further. 80 * The intuition is to equalize the time that is spend profiling each method. 81 * The same predicate is used to control the transition from level 3 to level 4 (C2). It should be 82 * noted though that the thresholds are relative. Moreover i and b for the 0->3 transition come 83 * from methodOop and for 3->4 transition they come from MDO (since profiled invocations are 84 * counted separately). 85 * 86 * OSR transitions are controlled simply with b > TierXBackEdgeThreshold * s predicates. 87 * 88 * - Tier?LoadFeedback options are used to automatically scale the predicates described above depending 89 * on the compiler load. The scaling coefficients are computed as follows: 90 * 91 * s = queue_size_X / (TierXLoadFeedback * compiler_count_X) + 1, 92 * 93 * where queue_size_X is the current size of the compiler queue of level X, and compiler_count_X 94 * is the number of level X compiler threads. 95 * 96 * Basically these parameters describe how many methods should be in the compile queue 97 * per compiler thread before the scaling coefficient increases by one. 98 * 99 * This feedback provides the mechanism to automatically control the flow of compilation requests 100 * depending on the machine speed, mutator load and other external factors. 101 * 102 * - Tier3DelayOn and Tier3DelayOff parameters control another important feedback loop. 103 * Consider the following observation: a method compiled with full profiling (level 3) 104 * is about 30% slower than a method at level 2 (just invocation and backedge counters, no MDO). 105 * Normally, the following transitions will occur: 0->3->4. The problem arises when the C2 queue 106 * gets congested and the 3->4 transition is delayed. While the method is the C2 queue it continues 107 * executing at level 3 for much longer time than is required by the predicate and at suboptimal speed. 108 * The idea is to dynamically change the behavior of the system in such a way that if a substantial 109 * load on C2 is detected we would first do the 0->2 transition allowing a method to run faster. 110 * And then when the load decreases to allow 2->3 transitions. 111 * 112 * Tier3Delay* parameters control this switching mechanism. 113 * Tier3DelayOn is the number of methods in the C2 queue per compiler thread after which the policy 114 * no longer does 0->3 transitions but does 0->2 transitions instead. 115 * Tier3DelayOff switches the original behavior back when the number of methods in the C2 queue 116 * per compiler thread falls below the specified amount. 117 * The hysteresis is necessary to avoid jitter. 118 * 119 * - TieredCompileTaskTimeout is the amount of time an idle method can spend in the compile queue. 120 * Basically, since we use the event rate d(i + b)/dt as a value of priority when selecting a method to 121 * compile from the compile queue, we also can detect stale methods for which the rate has been 122 * 0 for some time in the same iteration. Stale methods can appear in the queue when an application 123 * abruptly changes its behavior. 124 * 125 * - TieredStopAtLevel, is used mostly for testing. It allows to bypass the policy logic and stick 126 * to a given level. For example it's useful to set TieredStopAtLevel = 1 in order to compile everything 127 * with pure c1. 128 * 129 * - Tier0ProfilingStartPercentage allows the interpreter to start profiling when the inequalities in the 130 * 0->3 predicate are already exceeded by the given percentage but the level 3 version of the 131 * method is still not ready. We can even go directly from level 0 to 4 if c1 doesn't produce a compiled 132 * version in time. This reduces the overall transition to level 4 and decreases the startup time. 133 * Note that this behavior is also guarded by the Tier3Delay mechanism: when the c2 queue is too long 134 * these is not reason to start profiling prematurely. 135 * 136 * - TieredRateUpdateMinTime and TieredRateUpdateMaxTime are parameters of the rate computation. 137 * Basically, the rate is not computed more frequently than TieredRateUpdateMinTime and is considered 138 * to be zero if no events occurred in TieredRateUpdateMaxTime. 139 */ 140 141 142 class AdvancedThresholdPolicy : public SimpleThresholdPolicy { 143 jlong _start_time; 144 145 // Call and loop predicates determine whether a transition to a higher compilation 146 // level should be performed (pointers to predicate functions are passed to common(). 147 // Predicates also take compiler load into account. 148 typedef bool (AdvancedThresholdPolicy::*Predicate)(int i, int b, CompLevel cur_level); 149 bool call_predicate(int i, int b, CompLevel cur_level); 150 bool loop_predicate(int i, int b, CompLevel cur_level); 151 // Common transition function. Given a predicate determines if a method should transition to another level. 152 CompLevel common(Predicate p, methodOop method, CompLevel cur_level); 153 // Transition functions. 154 // call_event determines if a method should be compiled at a different 155 // level with a regular invocation entry. 156 CompLevel call_event(methodOop method, CompLevel cur_level); 157 // loop_event checks if a method should be OSR compiled at a different 158 // level. 159 CompLevel loop_event(methodOop method, CompLevel cur_level); 160 // Has a method been long around? 161 // We don't remove old methods from the compile queue even if they have 162 // very low activity (see select_task()). 163 inline bool is_old(methodOop method); 164 // Was a given method inactive for a given number of milliseconds. 165 // If it is, we would remove it from the queue (see select_task()). 166 inline bool is_stale(jlong t, jlong timeout, methodOop m); 167 // Compute the weight of the method for the compilation scheduling 168 inline double weight(methodOop method); 169 // Apply heuristics and return true if x should be compiled before y 170 inline bool compare_methods(methodOop x, methodOop y); 171 // Compute event rate for a given method. The rate is the number of event (invocations + backedges) 172 // per millisecond. 173 inline void update_rate(jlong t, methodOop m); 174 // Compute threshold scaling coefficient 175 inline double threshold_scale(CompLevel level, int feedback_k); 176 // If a method is old enough and is still in the interpreter we would want to 177 // start profiling without waiting for the compiled method to arrive. This function 178 // determines whether we should do that. 179 inline bool should_create_mdo(methodOop method, CompLevel cur_level); 180 // Create MDO if necessary. 181 void create_mdo(methodHandle mh, TRAPS); 182 // Is method profiled enough? 183 bool is_method_profiled(methodOop method); 184 185 protected: 186 void print_specific(EventType type, methodHandle mh, methodHandle imh, int bci, CompLevel level); 187 188 void set_start_time(jlong t) { _start_time = t; } 189 jlong start_time() const { return _start_time; } 190 191 // Submit a given method for compilation (and update the rate). 192 virtual void submit_compile(methodHandle mh, int bci, CompLevel level, TRAPS); 193 // event() from SimpleThresholdPolicy would call these. 194 virtual void method_invocation_event(methodHandle method, methodHandle inlinee, 195 CompLevel level, TRAPS); 196 virtual void method_back_branch_event(methodHandle method, methodHandle inlinee, 197 int bci, CompLevel level, TRAPS); 198 public: 199 AdvancedThresholdPolicy() : _start_time(0) { } 200 // Select task is called by CompileBroker. We should return a task or NULL. 201 virtual CompileTask* select_task(CompileQueue* compile_queue); 202 virtual void initialize(); 203 }; 204 205 #endif // TIERED 206 207 #endif // SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP