1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/src/share/vm/runtime/advancedThresholdPolicy.hpp Wed Apr 27 01:25:04 2016 +0800
1.3 @@ -0,0 +1,231 @@
1.4 +/*
1.5 + * Copyright (c) 2010, 2013, Oracle and/or its affiliates. All rights reserved.
1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
1.7 + *
1.8 + * This code is free software; you can redistribute it and/or modify it
1.9 + * under the terms of the GNU General Public License version 2 only, as
1.10 + * published by the Free Software Foundation.
1.11 + *
1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT
1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
1.15 + * version 2 for more details (a copy is included in the LICENSE file that
1.16 + * accompanied this code).
1.17 + *
1.18 + * You should have received a copy of the GNU General Public License version
1.19 + * 2 along with this work; if not, write to the Free Software Foundation,
1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
1.21 + *
1.22 + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
1.23 + * or visit www.oracle.com if you need additional information or have any
1.24 + * questions.
1.25 + *
1.26 + */
1.27 +
1.28 +#ifndef SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP
1.29 +#define SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP
1.30 +
1.31 +#include "runtime/simpleThresholdPolicy.hpp"
1.32 +
1.33 +#ifdef TIERED
1.34 +class CompileTask;
1.35 +class CompileQueue;
1.36 +
1.37 +/*
1.38 + * The system supports 5 execution levels:
1.39 + * * level 0 - interpreter
1.40 + * * level 1 - C1 with full optimization (no profiling)
1.41 + * * level 2 - C1 with invocation and backedge counters
1.42 + * * level 3 - C1 with full profiling (level 2 + MDO)
1.43 + * * level 4 - C2
1.44 + *
1.45 + * Levels 0, 2 and 3 periodically notify the runtime about the current value of the counters
1.46 + * (invocation counters and backedge counters). The frequency of these notifications is
1.47 + * different at each level. These notifications are used by the policy to decide what transition
1.48 + * to make.
1.49 + *
1.50 + * Execution starts at level 0 (interpreter), then the policy can decide either to compile the
1.51 + * method at level 3 or level 2. The decision is based on the following factors:
1.52 + * 1. The length of the C2 queue determines the next level. The observation is that level 2
1.53 + * is generally faster than level 3 by about 30%, therefore we would want to minimize the time
1.54 + * a method spends at level 3. We should only spend the time at level 3 that is necessary to get
1.55 + * adequate profiling. So, if the C2 queue is long enough it is more beneficial to go first to
1.56 + * level 2, because if we transitioned to level 3 we would be stuck there until our C2 compile
1.57 + * request makes its way through the long queue. When the load on C2 recedes we are going to
1.58 + * recompile at level 3 and start gathering profiling information.
1.59 + * 2. The length of C1 queue is used to dynamically adjust the thresholds, so as to introduce
1.60 + * additional filtering if the compiler is overloaded. The rationale is that by the time a
1.61 + * method gets compiled it can become unused, so it doesn't make sense to put too much onto the
1.62 + * queue.
1.63 + *
1.64 + * After profiling is completed at level 3 the transition is made to level 4. Again, the length
1.65 + * of the C2 queue is used as a feedback to adjust the thresholds.
1.66 + *
1.67 + * After the first C1 compile some basic information is determined about the code like the number
1.68 + * of the blocks and the number of the loops. Based on that it can be decided that a method
1.69 + * is trivial and compiling it with C1 will yield the same code. In this case the method is
1.70 + * compiled at level 1 instead of 4.
1.71 + *
1.72 + * We also support profiling at level 0. If C1 is slow enough to produce the level 3 version of
1.73 + * the code and the C2 queue is sufficiently small we can decide to start profiling in the
1.74 + * interpreter (and continue profiling in the compiled code once the level 3 version arrives).
1.75 + * If the profiling at level 0 is fully completed before level 3 version is produced, a level 2
1.76 + * version is compiled instead in order to run faster waiting for a level 4 version.
1.77 + *
1.78 + * Compile queues are implemented as priority queues - for each method in the queue we compute
1.79 + * the event rate (the number of invocation and backedge counter increments per unit of time).
1.80 + * When getting an element off the queue we pick the one with the largest rate. Maintaining the
1.81 + * rate also allows us to remove stale methods (the ones that got on the queue but stopped
1.82 + * being used shortly after that).
1.83 +*/
1.84 +
1.85 +/* Command line options:
1.86 + * - Tier?InvokeNotifyFreqLog and Tier?BackedgeNotifyFreqLog control the frequency of method
1.87 + * invocation and backedge notifications. Basically every n-th invocation or backedge a mutator thread
1.88 + * makes a call into the runtime.
1.89 + *
1.90 + * - Tier?CompileThreshold, Tier?BackEdgeThreshold, Tier?MinInvocationThreshold control
1.91 + * compilation thresholds.
1.92 + * Level 2 thresholds are not used and are provided for option-compatibility and potential future use.
1.93 + * Other thresholds work as follows:
1.94 + *
1.95 + * Transition from interpreter (level 0) to C1 with full profiling (level 3) happens when
1.96 + * the following predicate is true (X is the level):
1.97 + *
1.98 + * i > TierXInvocationThreshold * s || (i > TierXMinInvocationThreshold * s && i + b > TierXCompileThreshold * s),
1.99 + *
1.100 + * where $i$ is the number of method invocations, $b$ number of backedges and $s$ is the scaling
1.101 + * coefficient that will be discussed further.
1.102 + * The intuition is to equalize the time that is spend profiling each method.
1.103 + * The same predicate is used to control the transition from level 3 to level 4 (C2). It should be
1.104 + * noted though that the thresholds are relative. Moreover i and b for the 0->3 transition come
1.105 + * from Method* and for 3->4 transition they come from MDO (since profiled invocations are
1.106 + * counted separately).
1.107 + *
1.108 + * OSR transitions are controlled simply with b > TierXBackEdgeThreshold * s predicates.
1.109 + *
1.110 + * - Tier?LoadFeedback options are used to automatically scale the predicates described above depending
1.111 + * on the compiler load. The scaling coefficients are computed as follows:
1.112 + *
1.113 + * s = queue_size_X / (TierXLoadFeedback * compiler_count_X) + 1,
1.114 + *
1.115 + * where queue_size_X is the current size of the compiler queue of level X, and compiler_count_X
1.116 + * is the number of level X compiler threads.
1.117 + *
1.118 + * Basically these parameters describe how many methods should be in the compile queue
1.119 + * per compiler thread before the scaling coefficient increases by one.
1.120 + *
1.121 + * This feedback provides the mechanism to automatically control the flow of compilation requests
1.122 + * depending on the machine speed, mutator load and other external factors.
1.123 + *
1.124 + * - Tier3DelayOn and Tier3DelayOff parameters control another important feedback loop.
1.125 + * Consider the following observation: a method compiled with full profiling (level 3)
1.126 + * is about 30% slower than a method at level 2 (just invocation and backedge counters, no MDO).
1.127 + * Normally, the following transitions will occur: 0->3->4. The problem arises when the C2 queue
1.128 + * gets congested and the 3->4 transition is delayed. While the method is the C2 queue it continues
1.129 + * executing at level 3 for much longer time than is required by the predicate and at suboptimal speed.
1.130 + * The idea is to dynamically change the behavior of the system in such a way that if a substantial
1.131 + * load on C2 is detected we would first do the 0->2 transition allowing a method to run faster.
1.132 + * And then when the load decreases to allow 2->3 transitions.
1.133 + *
1.134 + * Tier3Delay* parameters control this switching mechanism.
1.135 + * Tier3DelayOn is the number of methods in the C2 queue per compiler thread after which the policy
1.136 + * no longer does 0->3 transitions but does 0->2 transitions instead.
1.137 + * Tier3DelayOff switches the original behavior back when the number of methods in the C2 queue
1.138 + * per compiler thread falls below the specified amount.
1.139 + * The hysteresis is necessary to avoid jitter.
1.140 + *
1.141 + * - TieredCompileTaskTimeout is the amount of time an idle method can spend in the compile queue.
1.142 + * Basically, since we use the event rate d(i + b)/dt as a value of priority when selecting a method to
1.143 + * compile from the compile queue, we also can detect stale methods for which the rate has been
1.144 + * 0 for some time in the same iteration. Stale methods can appear in the queue when an application
1.145 + * abruptly changes its behavior.
1.146 + *
1.147 + * - TieredStopAtLevel, is used mostly for testing. It allows to bypass the policy logic and stick
1.148 + * to a given level. For example it's useful to set TieredStopAtLevel = 1 in order to compile everything
1.149 + * with pure c1.
1.150 + *
1.151 + * - Tier0ProfilingStartPercentage allows the interpreter to start profiling when the inequalities in the
1.152 + * 0->3 predicate are already exceeded by the given percentage but the level 3 version of the
1.153 + * method is still not ready. We can even go directly from level 0 to 4 if c1 doesn't produce a compiled
1.154 + * version in time. This reduces the overall transition to level 4 and decreases the startup time.
1.155 + * Note that this behavior is also guarded by the Tier3Delay mechanism: when the c2 queue is too long
1.156 + * these is not reason to start profiling prematurely.
1.157 + *
1.158 + * - TieredRateUpdateMinTime and TieredRateUpdateMaxTime are parameters of the rate computation.
1.159 + * Basically, the rate is not computed more frequently than TieredRateUpdateMinTime and is considered
1.160 + * to be zero if no events occurred in TieredRateUpdateMaxTime.
1.161 + */
1.162 +
1.163 +
1.164 +class AdvancedThresholdPolicy : public SimpleThresholdPolicy {
1.165 + jlong _start_time;
1.166 +
1.167 + // Call and loop predicates determine whether a transition to a higher compilation
1.168 + // level should be performed (pointers to predicate functions are passed to common().
1.169 + // Predicates also take compiler load into account.
1.170 + typedef bool (AdvancedThresholdPolicy::*Predicate)(int i, int b, CompLevel cur_level);
1.171 + bool call_predicate(int i, int b, CompLevel cur_level);
1.172 + bool loop_predicate(int i, int b, CompLevel cur_level);
1.173 + // Common transition function. Given a predicate determines if a method should transition to another level.
1.174 + CompLevel common(Predicate p, Method* method, CompLevel cur_level, bool disable_feedback = false);
1.175 + // Transition functions.
1.176 + // call_event determines if a method should be compiled at a different
1.177 + // level with a regular invocation entry.
1.178 + CompLevel call_event(Method* method, CompLevel cur_level);
1.179 + // loop_event checks if a method should be OSR compiled at a different
1.180 + // level.
1.181 + CompLevel loop_event(Method* method, CompLevel cur_level);
1.182 + // Has a method been long around?
1.183 + // We don't remove old methods from the compile queue even if they have
1.184 + // very low activity (see select_task()).
1.185 + inline bool is_old(Method* method);
1.186 + // Was a given method inactive for a given number of milliseconds.
1.187 + // If it is, we would remove it from the queue (see select_task()).
1.188 + inline bool is_stale(jlong t, jlong timeout, Method* m);
1.189 + // Compute the weight of the method for the compilation scheduling
1.190 + inline double weight(Method* method);
1.191 + // Apply heuristics and return true if x should be compiled before y
1.192 + inline bool compare_methods(Method* x, Method* y);
1.193 + // Compute event rate for a given method. The rate is the number of event (invocations + backedges)
1.194 + // per millisecond.
1.195 + inline void update_rate(jlong t, Method* m);
1.196 + // Compute threshold scaling coefficient
1.197 + inline double threshold_scale(CompLevel level, int feedback_k);
1.198 + // If a method is old enough and is still in the interpreter we would want to
1.199 + // start profiling without waiting for the compiled method to arrive. This function
1.200 + // determines whether we should do that.
1.201 + inline bool should_create_mdo(Method* method, CompLevel cur_level);
1.202 + // Create MDO if necessary.
1.203 + void create_mdo(methodHandle mh, JavaThread* thread);
1.204 + // Is method profiled enough?
1.205 + bool is_method_profiled(Method* method);
1.206 +
1.207 + double _increase_threshold_at_ratio;
1.208 +
1.209 +protected:
1.210 + void print_specific(EventType type, methodHandle mh, methodHandle imh, int bci, CompLevel level);
1.211 +
1.212 + void set_increase_threshold_at_ratio() { _increase_threshold_at_ratio = 100 / (100 - (double)IncreaseFirstTierCompileThresholdAt); }
1.213 + void set_start_time(jlong t) { _start_time = t; }
1.214 + jlong start_time() const { return _start_time; }
1.215 +
1.216 + // Submit a given method for compilation (and update the rate).
1.217 + virtual void submit_compile(methodHandle mh, int bci, CompLevel level, JavaThread* thread);
1.218 + // event() from SimpleThresholdPolicy would call these.
1.219 + virtual void method_invocation_event(methodHandle method, methodHandle inlinee,
1.220 + CompLevel level, nmethod* nm, JavaThread* thread);
1.221 + virtual void method_back_branch_event(methodHandle method, methodHandle inlinee,
1.222 + int bci, CompLevel level, nmethod* nm, JavaThread* thread);
1.223 +public:
1.224 + AdvancedThresholdPolicy() : _start_time(0) { }
1.225 + // Select task is called by CompileBroker. We should return a task or NULL.
1.226 + virtual CompileTask* select_task(CompileQueue* compile_queue);
1.227 + virtual void initialize();
1.228 + virtual bool should_not_inline(ciEnv* env, ciMethod* callee);
1.229 +
1.230 +};
1.231 +
1.232 +#endif // TIERED
1.233 +
1.234 +#endif // SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP
