Cache替换策略
LRU, Least Recently Used, LRU算法根据各block(cache line)使用的情况, 总是选择那个最长时间未被使用的block进行替换。这种策略比较好的反映了程序局部性规律。
gem5中该替换策略的代码:
void LRURP::invalidate(const std::shared_ptr<ReplacementData>& replacement_data) const { // Reset last touch timestamp std::static_pointer_cast<LRUReplData>( replacement_data)->lastTouchTick = Tick(0); } void LRURP::touch(const std::shared_ptr<ReplacementData>& replacement_data) const { // Update last touch timestamp std::static_pointer_cast<LRUReplData>( replacement_data)->lastTouchTick = curTick(); } void LRURP::reset(const std::shared_ptr<ReplacementData>& replacement_data) const { // Set last touch timestamp std::static_pointer_cast<LRUReplData>( replacement_data)->lastTouchTick = curTick(); } ReplaceableEntry* LRURP::getVictim(const ReplacementCandidates& candidates) const { // There must be at least one replacement candidate assert(candidates.size() > 0); // Visit all candidates to find victim ReplaceableEntry* victim = candidates[0]; for (const auto& candidate : candidates) { // Update victim entry if necessary if (std::static_pointer_cast<LRUReplData>( candidate->replacementData)->lastTouchTick < std::static_pointer_cast<LRUReplData>( victim->replacementData)->lastTouchTick) { victim = candidate; } } return victim; }
MRU(Most Recently Used)和LRU类似,差别在于选择最近被使用的block进行替换。
gem5中该替换策略的代码:
void MRURP::invalidate(const std::shared_ptr<ReplacementData>& replacement_data) const { // Reset last touch timestamp std::static_pointer_cast<MRUReplData>( replacement_data)->lastTouchTick = Tick(0); } void MRURP::touch(const std::shared_ptr<ReplacementData>& replacement_data) const { // Update last touch timestamp std::static_pointer_cast<MRUReplData>( replacement_data)->lastTouchTick = curTick(); } void MRURP::reset(const std::shared_ptr<ReplacementData>& replacement_data) const { // Set last touch timestamp std::static_pointer_cast<MRUReplData>( replacement_data)->lastTouchTick = curTick(); } ReplaceableEntry* MRURP::getVictim(const ReplacementCandidates& candidates) const { // There must be at least one replacement candidate assert(candidates.size() > 0); // Visit all candidates to find victim ReplaceableEntry* victim = candidates[0]; for (const auto& candidate : candidates) { std::shared_ptr<MRUReplData> candidate_replacement_data = std::static_pointer_cast<MRUReplData>(candidate->replacementData); // Stop searching entry if a cache line that doesn't warm up is found. if (candidate_replacement_data->lastTouchTick == 0) { victim = candidate; break; } else if (candidate_replacement_data->lastTouchTick > std::static_pointer_cast<MRUReplData>( victim->replacementData)->lastTouchTick) { victim = candidate; } } return victim; }
Random,随机选择一个block进行替换。
gem5中该替换策略的代码:
void RandomRP::invalidate(const std::shared_ptr<ReplacementData>& replacement_data) const { // Unprioritize replacement data victimization std::static_pointer_cast<RandomReplData>( replacement_data)->valid = false; } void RandomRP::touch(const std::shared_ptr<ReplacementData>& replacement_data) const { } void RandomRP::reset(const std::shared_ptr<ReplacementData>& replacement_data) const { // Unprioritize replacement data victimization std::static_pointer_cast<RandomReplData>( replacement_data)->valid = true; } ReplaceableEntry* RandomRP::getVictim(const ReplacementCandidates& candidates) const { // There must be at least one replacement candidate assert(candidates.size() > 0); // Choose one candidate at random ReplaceableEntry* victim = candidates[random_mt.random<unsigned>(0, candidates.size() - 1)]; // Visit all candidates to search for an invalid entry. If one is found, // its eviction is prioritized for (const auto& candidate : candidates) { if (!std::static_pointer_cast<RandomReplData>( candidate->replacementData)->valid) { victim = candidate; break; } }
LFU(Least Frequently Used),最近最少被使用次数的block被替换,每个block都有一个引用计数,每次替换该block,都会对该计数加1。
gem5中该替换策略的代码:
void LFURP::invalidate(const std::shared_ptr<ReplacementData>& replacement_data) const { // Reset reference count std::static_pointer_cast<LFUReplData>(replacement_data)->refCount = 0; } void LFURP::touch(const std::shared_ptr<ReplacementData>& replacement_data) const { // Update reference count std::static_pointer_cast<LFUReplData>(replacement_data)->refCount++; } void LFURP::reset(const std::shared_ptr<ReplacementData>& replacement_data) const { // Reset reference count std::static_pointer_cast<LFUReplData>(replacement_data)->refCount = 1; } ReplaceableEntry* LFURP::getVictim(const ReplacementCandidates& candidates) const { // There must be at least one replacement candidate assert(candidates.size() > 0); // Visit all candidates to find victim ReplaceableEntry* victim = candidates[0]; for (const auto& candidate : candidates) { // Update victim entry if necessary if (std::static_pointer_cast<LFUReplData>( candidate->replacementData)->refCount < std::static_pointer_cast<LFUReplData>( victim->replacementData)->refCount) { victim = candidate; } } return victim; }
FIFO(First in First out), 最先使用过的block,最先被替换。
gem5中该替换策略的代码:
void FIFORP::invalidate(const std::shared_ptr<ReplacementData>& replacement_data) const { // Reset insertion tick std::static_pointer_cast<FIFOReplData>( replacement_data)->tickInserted = Tick(0); } void FIFORP::touch(const std::shared_ptr<ReplacementData>& replacement_data) const { // A touch does not modify the insertion tick } void FIFORP::reset(const std::shared_ptr<ReplacementData>& replacement_data) const { // Set insertion tick std::static_pointer_cast<FIFOReplData>( replacement_data)->tickInserted = curTick(); } ReplaceableEntry* FIFORP::getVictim(const ReplacementCandidates& candidates) const { // There must be at least one replacement candidate assert(candidates.size() > 0); // Visit all candidates to find victim ReplaceableEntry* victim = candidates[0]; for (const auto& candidate : candidates) { // Update victim entry if necessary if (std::static_pointer_cast<FIFOReplData>( candidate->replacementData)->tickInserted < std::static_pointer_cast<FIFOReplData>( victim->replacementData)->tickInserted) { victim = candidate; } } return victim; }
BIP,(Bimodal Insertion Policy)替换策略,是LRU和MRU的结合体,大概率采用MRU替换,小概率采用LRU策略。
gem5中该替换策略的代码
struct BIPRPParams; class BIPRP : public LRURP { protected: /** * Bimodal throtle parameter. Value in the range [0,100] used to decide * if a new entry is inserted at the MRU or LRU position. */ const unsigned btp; public: /** Convenience typedef. */ typedef BIPRPParams Params; /** * Construct and initiliaze this replacement policy. */ BIPRP(const Params *p); /** * Destructor. */ ~BIPRP() {} /** * Reset replacement data for an entry. Used when an entry is inserted. * Uses the bimodal throtle parameter to decide whether the new entry * should be inserted as MRU, or LRU. * * @param replacement_data Replacement data to be reset. */ void reset(const std::shared_ptr<ReplacementData>& replacement_data) const override; }; void BIPRP::reset(const std::shared_ptr<ReplacementData>& replacement_data) const { std::shared_ptr<LRUReplData> casted_replacement_data = std::static_pointer_cast<LRUReplData>(replacement_data); // Entries are inserted as MRU if lower than btp, LRU otherwise if (random_mt.random<unsigned>(1, 100) <= btp) { casted_replacement_data->lastTouchTick = curTick(); } else { // Make their timestamps as old as possible, so that they become LRU casted_replacement_data->lastTouchTick = 1; } }
NRU(Not Recent Used) 是LRU的一个近似策略,被广泛应用于现代高性能处理器中。应用NRU策略的cache,需要在每个cache block中增加一位标记,该标记(NRU bit)“0”表示最近可能被访问到的,“1”表示最近不能访问到的。
每当一个cache hit,该cache block的NRU bit被设置为“0”表示在最近的将来,该cache block很有可能再被访问到;每当一个cache miss,替换算法会从左至右扫描NRU bit为“1”的block,如果找到则替换出该cache block,并将新插入的cache block 的NRU bit置为“0”,如果没有找到,那么将所有cache block的NRU bit置为“1”,重新从左至右扫描。
STATIC RRIP, 该替换策略是对NRU的扩展,其将NRU bit扩展成M位,当M=1时,该算法蜕化成NRU。而扩展成M位的原因是为了更细粒度的区分cache block,而不是只有两个状态(最近将要访问和最近最远将要访问)。
该算法的描述和NRU相同,每当一个cache hit,该cache block的NRU bit被设置为“0”表示在最近的将来,该cache block很有可能再被访问到;每当一个cache miss,替换算法会从左至右扫描NRU bit为“2^M -1”的block,如果找到则替换出该cache block,并将新插入的cache block 的NRU bit置为“2^M -2”,如果没有找到,那么将所有cache block的NRU bit增加1,重新从左至右扫描。
上面将新插入的cache block设置为“2^M -2”,主要是为了防止那些很久才能被再次使用到的cache block长期占用cache空间, 但这样确实会影响那些空间局部性很好的程序的性能。
在RRIP类的策略中,NRU bit被描述为RRPV(Re- reference Prediction Values),可以理解为当前block被替换出去的可能性,越高越容易被替换出去。
DYNAMIC RRIP(Bimodal RRIP), 对Static RRIP来讲,如果程序的工作集大于cache容量,那么将会频繁的换进换出,造成抖动。为此,Bimodal RRIP提出,对于新插入的cache block,以较大概率设置NRU bits为“2^M -1",同时以较小概率设置为”2^M -2",一次来避免抖动。
那么对于混合的访存序列,应该使用SRRIP还是BRRIP的问题,一种称之为“set Dueling”的技术将两种技术应用到不同的两个cache set上,然后统计两个set上的运行情况(主要是命中率),然后来决断到底使用两种技术中的哪一个,然后将该算法策略部署到其余各个set上。
GEM5中也有BRRIP替换策略的实现。