Mercurial > hg > graal-compiler
diff src/share/vm/runtime/advancedThresholdPolicy.hpp @ 2348:5d8f5a6dced7
7020403: Add AdvancedCompilationPolicy for tiered
Summary: This implements adaptive tiered compilation policy.
Reviewed-by: kvn, never
author | iveresov |
---|---|
date | Fri, 04 Mar 2011 15:14:16 -0800 |
parents | |
children | 97b64f73103b |
line wrap: on
line diff
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/share/vm/runtime/advancedThresholdPolicy.hpp Fri Mar 04 15:14:16 2011 -0800 @@ -0,0 +1,207 @@ +/* +* Copyright (c) 2010, 2011 Oracle and/or its affiliates. All rights reserved. +* ORACLE PROPRIETARY/CONFIDENTIAL. Use is subject to license terms. +*/ + +#ifndef SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP +#define SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP + +#include "runtime/simpleThresholdPolicy.hpp" + +#ifdef TIERED +class CompileTask; +class CompileQueue; + +/* + * The system supports 5 execution levels: + * * level 0 - interpreter + * * level 1 - C1 with full optimization (no profiling) + * * level 2 - C1 with invocation and backedge counters + * * level 3 - C1 with full profiling (level 2 + MDO) + * * level 4 - C2 + * + * Levels 0, 2 and 3 periodically notify the runtime about the current value of the counters + * (invocation counters and backedge counters). The frequency of these notifications is + * different at each level. These notifications are used by the policy to decide what transition + * to make. + * + * Execution starts at level 0 (interpreter), then the policy can decide either to compile the + * method at level 3 or level 2. The decision is based on the following factors: + * 1. The length of the C2 queue determines the next level. The observation is that level 2 + * is generally faster than level 3 by about 30%, therefore we would want to minimize the time + * a method spends at level 3. We should only spend the time at level 3 that is necessary to get + * adequate profiling. So, if the C2 queue is long enough it is more beneficial to go first to + * level 2, because if we transitioned to level 3 we would be stuck there until our C2 compile + * request makes its way through the long queue. When the load on C2 recedes we are going to + * recompile at level 3 and start gathering profiling information. + * 2. The length of C1 queue is used to dynamically adjust the thresholds, so as to introduce + * additional filtering if the compiler is overloaded. The rationale is that by the time a + * method gets compiled it can become unused, so it doesn't make sense to put too much onto the + * queue. + * + * After profiling is completed at level 3 the transition is made to level 4. Again, the length + * of the C2 queue is used as a feedback to adjust the thresholds. + * + * After the first C1 compile some basic information is determined about the code like the number + * of the blocks and the number of the loops. Based on that it can be decided that a method + * is trivial and compiling it with C1 will yield the same code. In this case the method is + * compiled at level 1 instead of 4. + * + * We also support profiling at level 0. If C1 is slow enough to produce the level 3 version of + * the code and the C2 queue is sufficiently small we can decide to start profiling in the + * interpreter (and continue profiling in the compiled code once the level 3 version arrives). + * If the profiling at level 0 is fully completed before level 3 version is produced, a level 2 + * version is compiled instead in order to run faster waiting for a level 4 version. + * + * Compile queues are implemented as priority queues - for each method in the queue we compute + * the event rate (the number of invocation and backedge counter increments per unit of time). + * When getting an element off the queue we pick the one with the largest rate. Maintaining the + * rate also allows us to remove stale methods (the ones that got on the queue but stopped + * being used shortly after that). +*/ + +/* Command line options: + * - Tier?InvokeNotifyFreqLog and Tier?BackedgeNotifyFreqLog control the frequency of method + * invocation and backedge notifications. Basically every n-th invocation or backedge a mutator thread + * makes a call into the runtime. + * + * - Tier?CompileThreshold, Tier?BackEdgeThreshold, Tier?MinInvocationThreshold control + * compilation thresholds. + * Level 2 thresholds are not used and are provided for option-compatibility and potential future use. + * Other thresholds work as follows: + * + * Transition from interpreter (level 0) to C1 with full profiling (level 3) happens when + * the following predicate is true (X is the level): + * + * i > TierXInvocationThreshold * s || (i > TierXMinInvocationThreshold * s && i + b > TierXCompileThreshold * s), + * + * where $i$ is the number of method invocations, $b$ number of backedges and $s$ is the scaling + * coefficient that will be discussed further. + * The intuition is to equalize the time that is spend profiling each method. + * The same predicate is used to control the transition from level 3 to level 4 (C2). It should be + * noted though that the thresholds are relative. Moreover i and b for the 0->3 transition come + * from methodOop and for 3->4 transition they come from MDO (since profiled invocations are + * counted separately). + * + * OSR transitions are controlled simply with b > TierXBackEdgeThreshold * s predicates. + * + * - Tier?LoadFeedback options are used to automatically scale the predicates described above depending + * on the compiler load. The scaling coefficients are computed as follows: + * + * s = queue_size_X / (TierXLoadFeedback * compiler_count_X) + 1, + * + * where queue_size_X is the current size of the compiler queue of level X, and compiler_count_X + * is the number of level X compiler threads. + * + * Basically these parameters describe how many methods should be in the compile queue + * per compiler thread before the scaling coefficient increases by one. + * + * This feedback provides the mechanism to automatically control the flow of compilation requests + * depending on the machine speed, mutator load and other external factors. + * + * - Tier3DelayOn and Tier3DelayOff parameters control another important feedback loop. + * Consider the following observation: a method compiled with full profiling (level 3) + * is about 30% slower than a method at level 2 (just invocation and backedge counters, no MDO). + * Normally, the following transitions will occur: 0->3->4. The problem arises when the C2 queue + * gets congested and the 3->4 transition is delayed. While the method is the C2 queue it continues + * executing at level 3 for much longer time than is required by the predicate and at suboptimal speed. + * The idea is to dynamically change the behavior of the system in such a way that if a substantial + * load on C2 is detected we would first do the 0->2 transition allowing a method to run faster. + * And then when the load decreases to allow 2->3 transitions. + * + * Tier3Delay* parameters control this switching mechanism. + * Tier3DelayOn is the number of methods in the C2 queue per compiler thread after which the policy + * no longer does 0->3 transitions but does 0->2 transitions instead. + * Tier3DelayOff switches the original behavior back when the number of methods in the C2 queue + * per compiler thread falls below the specified amount. + * The hysteresis is necessary to avoid jitter. + * + * - TieredCompileTaskTimeout is the amount of time an idle method can spend in the compile queue. + * Basically, since we use the event rate d(i + b)/dt as a value of priority when selecting a method to + * compile from the compile queue, we also can detect stale methods for which the rate has been + * 0 for some time in the same iteration. Stale methods can appear in the queue when an application + * abruptly changes its behavior. + * + * - TieredStopAtLevel, is used mostly for testing. It allows to bypass the policy logic and stick + * to a given level. For example it's useful to set TieredStopAtLevel = 1 in order to compile everything + * with pure c1. + * + * - Tier0ProfilingStartPercentage allows the interpreter to start profiling when the inequalities in the + * 0->3 predicate are already exceeded by the given percentage but the level 3 version of the + * method is still not ready. We can even go directly from level 0 to 4 if c1 doesn't produce a compiled + * version in time. This reduces the overall transition to level 4 and decreases the startup time. + * Note that this behavior is also guarded by the Tier3Delay mechanism: when the c2 queue is too long + * these is not reason to start profiling prematurely. + * + * - TieredRateUpdateMinTime and TieredRateUpdateMaxTime are parameters of the rate computation. + * Basically, the rate is not computed more frequently than TieredRateUpdateMinTime and is considered + * to be zero if no events occurred in TieredRateUpdateMaxTime. + */ + + +class AdvancedThresholdPolicy : public SimpleThresholdPolicy { + jlong _start_time; + + // Call and loop predicates determine whether a transition to a higher compilation + // level should be performed (pointers to predicate functions are passed to common(). + // Predicates also take compiler load into account. + typedef bool (AdvancedThresholdPolicy::*Predicate)(int i, int b, CompLevel cur_level); + bool call_predicate(int i, int b, CompLevel cur_level); + bool loop_predicate(int i, int b, CompLevel cur_level); + // Common transition function. Given a predicate determines if a method should transition to another level. + CompLevel common(Predicate p, methodOop method, CompLevel cur_level); + // Transition functions. + // call_event determines if a method should be compiled at a different + // level with a regular invocation entry. + CompLevel call_event(methodOop method, CompLevel cur_level); + // loop_event checks if a method should be OSR compiled at a different + // level. + CompLevel loop_event(methodOop method, CompLevel cur_level); + // Has a method been long around? + // We don't remove old methods from the compile queue even if they have + // very low activity (see select_task()). + inline bool is_old(methodOop method); + // Was a given method inactive for a given number of milliseconds. + // If it is, we would remove it from the queue (see select_task()). + inline bool is_stale(jlong t, jlong timeout, methodOop m); + // Compute the weight of the method for the compilation scheduling + inline double weight(methodOop method); + // Apply heuristics and return true if x should be compiled before y + inline bool compare_methods(methodOop x, methodOop y); + // Compute event rate for a given method. The rate is the number of event (invocations + backedges) + // per millisecond. + inline void update_rate(jlong t, methodOop m); + // Compute threshold scaling coefficient + inline double threshold_scale(CompLevel level, int feedback_k); + // If a method is old enough and is still in the interpreter we would want to + // start profiling without waiting for the compiled method to arrive. This function + // determines whether we should do that. + inline bool should_create_mdo(methodOop method, CompLevel cur_level); + // Create MDO if necessary. + void create_mdo(methodHandle mh, TRAPS); + // Is method profiled enough? + bool is_method_profiled(methodOop method); + +protected: + void print_specific(EventType type, methodHandle mh, methodHandle imh, int bci, CompLevel level); + + void set_start_time(jlong t) { _start_time = t; } + jlong start_time() const { return _start_time; } + + // Submit a given method for compilation (and update the rate). + virtual void submit_compile(methodHandle mh, int bci, CompLevel level, TRAPS); + // event() from SimpleThresholdPolicy would call these. + virtual void method_invocation_event(methodHandle method, methodHandle inlinee, + CompLevel level, TRAPS); + virtual void method_back_branch_event(methodHandle method, methodHandle inlinee, + int bci, CompLevel level, TRAPS); +public: + AdvancedThresholdPolicy() : _start_time(0) { } + // Select task is called by CompileBroker. We should return a task or NULL. + virtual CompileTask* select_task(CompileQueue* compile_queue); + virtual void initialize(); +}; + +#endif // TIERED + +#endif // SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP