Top

AdvancedThresholdPolicy クラス


AdvancedThresholdPolicy

概要(Summary)

C1/C2 JIT Compiler 用の補助クラス (#ifdef TIERED 時にしか定義されない).

Tiered Compilation 用の CompilationPolicy クラス(= メソッドをコンパイルするかどうかの条件/閾値を管理するクラス) (See: here and here for details).

なお, このクラスは CompilationPolicyChoice オプションが 3 の場合にのみ使用される (そして Tiered Compilation 使用時のデフォルトは 3). (See: compilationPolicy_init(), Arguments::set_tiered_flags())

    ((cite: hotspot/src/share/vm/runtime/advancedThresholdPolicy.hpp))
    #ifdef TIERED
    ((cite: hotspot/src/share/vm/runtime/advancedThresholdPolicy.hpp))
    /*
     *  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).
    */
    ((cite: hotspot/src/share/vm/runtime/advancedThresholdPolicy.hpp))
    class AdvancedThresholdPolicy : public SimpleThresholdPolicy {

使われ方(Usage)

インスタンスの格納場所(where its instances are stored)

CompilationPolicy クラスの _policy フィールド (static フィールド) に(のみ)格納されている.

生成箇所(where its instances are created)

compilationPolicy_init() 内で(のみ)生成されている. そして, この関数は現在は以下のパスで(のみ)呼び出されている.

(HotSpot の起動時処理) (See: here for details)
-> Threads::create_vm()
   -> init_globals()
      -> compilationPolicy_init()

備考(Notes)

関連するコマンドラインオプションには以下のようなものがある.

    ((cite: hotspot/src/share/vm/runtime/advancedThresholdPolicy.hpp))
    /* 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.
     */

詳細(Details)

See: here for details



This document is available under the GNU GENERAL PUBLIC LICENSE Version 2.