projjal commented on a change in pull request #10465: URL: https://github.com/apache/arrow/pull/10465#discussion_r668062139
########## File path: cpp/src/gandiva/greedy_dual_size_cache.h ########## @@ -0,0 +1,153 @@ +// Licensed to the Apache Software Foundation (ASF) under one +// or more contributor license agreements. See the NOTICE file +// distributed with this work for additional information +// regarding copyright ownership. The ASF licenses this file +// to you under the Apache License, Version 2.0 (the +// "License"); you may not use this file except in compliance +// with the License. You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, +// software distributed under the License is distributed on an +// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +// KIND, either express or implied. See the License for the +// specific language governing permissions and limitations +// under the License. + +#pragma once + +#include <list> +#include <queue> +#include <set> +#include <unordered_map> +#include <utility> + +#include "arrow/util/optional.h" + +// modified cache to support evict policy using the GreedyDual-Size algorithm. +namespace gandiva { +// Defines a base value object supported on the cache that may contain properties +template <typename ValueType> +class ValueCacheObject { + public: + ValueCacheObject(ValueType module, uint64_t cost) : module(module), cost(cost) {} + ValueCacheObject() = default; Review comment: This is not required ########## File path: cpp/src/gandiva/projector.cc ########## @@ -186,7 +192,9 @@ Status Projector::Make(SchemaPtr schema, const ExpressionVector& exprs, // Instantiate the projector with the completely built llvm generator *projector = std::shared_ptr<Projector>( new Projector(std::move(llvm_gen), schema, output_fields, configuration)); - cache.PutModule(cache_key, *projector); + ValueCacheObject<std::shared_ptr<Projector>> value_cache = Review comment: Write it as: ``` ValueCacheObject<std::shared_ptr<Projector>> value_cache(*projector, elapsed); ``` ########## File path: cpp/src/gandiva/greedy_dual_size_cache.h ########## @@ -0,0 +1,153 @@ +// Licensed to the Apache Software Foundation (ASF) under one +// or more contributor license agreements. See the NOTICE file +// distributed with this work for additional information +// regarding copyright ownership. The ASF licenses this file +// to you under the Apache License, Version 2.0 (the +// "License"); you may not use this file except in compliance +// with the License. You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, +// software distributed under the License is distributed on an +// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +// KIND, either express or implied. See the License for the +// specific language governing permissions and limitations +// under the License. + +#pragma once + +#include <list> +#include <queue> +#include <set> +#include <unordered_map> +#include <utility> + +#include "arrow/util/optional.h" + +// modified cache to support evict policy using the GreedyDual-Size algorithm. +namespace gandiva { +// Defines a base value object supported on the cache that may contain properties +template <typename ValueType> +class ValueCacheObject { + public: + ValueCacheObject(ValueType module, uint64_t cost) : module(module), cost(cost) {} + ValueCacheObject() = default; + ValueType module; + uint64_t cost; + bool operator<(const ValueCacheObject& other) const { return cost < other.cost; } +}; + +// A particular cache based on the GreedyDual-Size cache which is a generalization of LRU +// which defines costs for each cache values. +// The algorithm associates a cost, C, with each cache value. Initially, when the value +// is brought into cache, C is set to be the cost related to the value (the cost is +// always non-negative). When a replacement needs to be made, the value with the lowest C +// cost is replaced, and then all values reduce their C costs by the minimum value of C +// over all the values already in the cache. +// If a value is accessed, its C value is restored to its initial cost. Thus, the C costs +// of recently accessed values retain a larger portion of the original cost than those of +// values that have not been accessed for a long time. The C costs are reduced as time +// goes and are restored when accessed. + +template <class Key, class Value> +class GreedyDualSizeCache { + // inner class to define the priority item + class PriorityItem { + public: + PriorityItem(uint64_t actual_priority, uint64_t original_priority, Key key) + : actual_priority(actual_priority), + original_priority(original_priority), + cache_key(key) {} + // this ensure that the items with low priority stays in the beginning of the queue, + // so it can be the one removed by evict operation + bool operator<(const PriorityItem& other) const { + return actual_priority < other.actual_priority; + } + uint64_t actual_priority; + uint64_t original_priority; + Key cache_key; + }; + + public: + struct hasher { + template <typename I> + std::size_t operator()(const I& i) const { + return i.Hash(); + } + }; + // a map from 'key' to a pair of Value and a pointer to the priority value + using map_type = std::unordered_map< + Key, std::pair<ValueCacheObject<Value>, typename std::set<PriorityItem>::iterator>, + hasher>; + + explicit GreedyDualSizeCache(size_t capacity) : inflation_(0), capacity_(capacity) {} + + ~GreedyDualSizeCache() = default; + + size_t size() const { return map_.size(); } + + size_t capacity() const { return capacity_; } + + bool empty() const { return map_.empty(); } + + bool contains(const Key& key) { return map_.find(key) != map_.end(); } + + void insert(const Key& key, const ValueCacheObject<Value>& value) { + typename map_type::iterator i = map_.find(key); + // check if element is not in the cache to add it + if (i == map_.end()) { + // insert item into the cache, but first check if it is full, to evict an item + // if it is necessary + if (size() >= capacity_) { + evict(); + } + + // insert the new item + auto iter = + priority_set_.insert(PriorityItem(value.cost + inflation_, value.cost, key)); + // save on map the value and the priority item iterator position + map_[key] = std::make_pair(value, iter.first); + } + } + + arrow::util::optional<ValueCacheObject<Value>> get(const Key& key) { + // lookup value in the cache + typename map_type::iterator value_for_key = map_.find(key); + if (value_for_key == map_.end()) { + // value not in cache + return arrow::util::nullopt; + } + PriorityItem item = *value_for_key->second.second; + // if the value was found on the cache, update its cost (original + inflation) + if (item.actual_priority != item.original_priority + inflation_) { + priority_set_.erase(value_for_key->second.second); + auto iter = priority_set_.insert(PriorityItem( + item.original_priority + inflation_, item.original_priority, item.cache_key)); + value_for_key->second.second = iter.first; + } + return value_for_key->second.first; + } + + void clear() { + map_.clear(); + priority_set_.clear(); + } + + private: + void evict() { + // evict item from the beginning of the set. This set is ordered from the + // lower priority value to the higher priority value. + typename std::set<PriorityItem>::iterator i = priority_set_.begin(); + // update the inflation cost related to the evicted item + inflation_ = (*i).actual_priority; + map_.erase((*i).cache_key); + priority_set_.erase(i); + } + + map_type map_; + std::set<PriorityItem> priority_set_; + uint64_t inflation_; Review comment: It is int64 so it can overflow for repeated inflation. Assuming 100ms avg compilation time, it is in the order of 10^11 which is large enough that we can ignore. But still I think we can measure the duration cost as microseconds/milliseconds instead of nano seconds. What is the compilation time for smallest possible expression `out = in` ? ########## File path: cpp/src/gandiva/greedy_dual_size_cache.h ########## @@ -0,0 +1,153 @@ +// Licensed to the Apache Software Foundation (ASF) under one +// or more contributor license agreements. See the NOTICE file +// distributed with this work for additional information +// regarding copyright ownership. The ASF licenses this file +// to you under the Apache License, Version 2.0 (the +// "License"); you may not use this file except in compliance +// with the License. You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, +// software distributed under the License is distributed on an +// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +// KIND, either express or implied. See the License for the +// specific language governing permissions and limitations +// under the License. + +#pragma once + +#include <list> +#include <queue> +#include <set> +#include <unordered_map> +#include <utility> + +#include "arrow/util/optional.h" + +// modified cache to support evict policy using the GreedyDual-Size algorithm. +namespace gandiva { +// Defines a base value object supported on the cache that may contain properties +template <typename ValueType> +class ValueCacheObject { + public: + ValueCacheObject(ValueType module, uint64_t cost) : module(module), cost(cost) {} + ValueCacheObject() = default; + ValueType module; + uint64_t cost; + bool operator<(const ValueCacheObject& other) const { return cost < other.cost; } +}; + +// A particular cache based on the GreedyDual-Size cache which is a generalization of LRU +// which defines costs for each cache values. +// The algorithm associates a cost, C, with each cache value. Initially, when the value +// is brought into cache, C is set to be the cost related to the value (the cost is +// always non-negative). When a replacement needs to be made, the value with the lowest C +// cost is replaced, and then all values reduce their C costs by the minimum value of C +// over all the values already in the cache. +// If a value is accessed, its C value is restored to its initial cost. Thus, the C costs +// of recently accessed values retain a larger portion of the original cost than those of +// values that have not been accessed for a long time. The C costs are reduced as time +// goes and are restored when accessed. + +template <class Key, class Value> +class GreedyDualSizeCache { + // inner class to define the priority item + class PriorityItem { + public: + PriorityItem(uint64_t actual_priority, uint64_t original_priority, Key key) + : actual_priority(actual_priority), + original_priority(original_priority), + cache_key(key) {} + // this ensure that the items with low priority stays in the beginning of the queue, + // so it can be the one removed by evict operation + bool operator<(const PriorityItem& other) const { + return actual_priority < other.actual_priority; + } + uint64_t actual_priority; + uint64_t original_priority; + Key cache_key; + }; + + public: + struct hasher { + template <typename I> + std::size_t operator()(const I& i) const { + return i.Hash(); + } + }; + // a map from 'key' to a pair of Value and a pointer to the priority value + using map_type = std::unordered_map< + Key, std::pair<ValueCacheObject<Value>, typename std::set<PriorityItem>::iterator>, + hasher>; + + explicit GreedyDualSizeCache(size_t capacity) : inflation_(0), capacity_(capacity) {} + + ~GreedyDualSizeCache() = default; + + size_t size() const { return map_.size(); } + + size_t capacity() const { return capacity_; } + + bool empty() const { return map_.empty(); } + + bool contains(const Key& key) { return map_.find(key) != map_.end(); } + + void insert(const Key& key, const ValueCacheObject<Value>& value) { + typename map_type::iterator i = map_.find(key); + // check if element is not in the cache to add it + if (i == map_.end()) { + // insert item into the cache, but first check if it is full, to evict an item + // if it is necessary + if (size() >= capacity_) { + evict(); + } + + // insert the new item + auto iter = + priority_set_.insert(PriorityItem(value.cost + inflation_, value.cost, key)); + // save on map the value and the priority item iterator position + map_[key] = std::make_pair(value, iter.first); + } + } + + arrow::util::optional<ValueCacheObject<Value>> get(const Key& key) { + // lookup value in the cache + typename map_type::iterator value_for_key = map_.find(key); + if (value_for_key == map_.end()) { + // value not in cache + return arrow::util::nullopt; + } + PriorityItem item = *value_for_key->second.second; + // if the value was found on the cache, update its cost (original + inflation) + if (item.actual_priority != item.original_priority + inflation_) { + priority_set_.erase(value_for_key->second.second); + auto iter = priority_set_.insert(PriorityItem( + item.original_priority + inflation_, item.original_priority, item.cache_key)); + value_for_key->second.second = iter.first; + } + return value_for_key->second.first; + } + + void clear() { + map_.clear(); + priority_set_.clear(); + } + + private: + void evict() { + // evict item from the beginning of the set. This set is ordered from the + // lower priority value to the higher priority value. + typename std::set<PriorityItem>::iterator i = priority_set_.begin(); + // update the inflation cost related to the evicted item + inflation_ = (*i).actual_priority; + map_.erase((*i).cache_key); + priority_set_.erase(i); + } + + map_type map_; + std::set<PriorityItem> priority_set_; + uint64_t inflation_; Review comment: It is int64 so it can overflow for repeated inflation. Assuming 100ms avg compilation time, it is in the order of 10^11 which I don't know if it is large enough that we can ignore. Its better to measure the duration cost as microseconds/milliseconds instead of nano seconds. What is the compilation time for smallest possible expression `out = in` ? ########## File path: cpp/src/gandiva/greedy_dual_size_cache.h ########## @@ -0,0 +1,153 @@ +// Licensed to the Apache Software Foundation (ASF) under one +// or more contributor license agreements. See the NOTICE file +// distributed with this work for additional information +// regarding copyright ownership. The ASF licenses this file +// to you under the Apache License, Version 2.0 (the +// "License"); you may not use this file except in compliance +// with the License. You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, +// software distributed under the License is distributed on an +// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +// KIND, either express or implied. See the License for the +// specific language governing permissions and limitations +// under the License. + +#pragma once + +#include <list> +#include <queue> +#include <set> +#include <unordered_map> +#include <utility> + +#include "arrow/util/optional.h" + +// modified cache to support evict policy using the GreedyDual-Size algorithm. +namespace gandiva { +// Defines a base value object supported on the cache that may contain properties +template <typename ValueType> +class ValueCacheObject { + public: + ValueCacheObject(ValueType module, uint64_t cost) : module(module), cost(cost) {} + ValueCacheObject() = default; + ValueType module; + uint64_t cost; + bool operator<(const ValueCacheObject& other) const { return cost < other.cost; } +}; + +// A particular cache based on the GreedyDual-Size cache which is a generalization of LRU +// which defines costs for each cache values. +// The algorithm associates a cost, C, with each cache value. Initially, when the value +// is brought into cache, C is set to be the cost related to the value (the cost is +// always non-negative). When a replacement needs to be made, the value with the lowest C +// cost is replaced, and then all values reduce their C costs by the minimum value of C +// over all the values already in the cache. +// If a value is accessed, its C value is restored to its initial cost. Thus, the C costs +// of recently accessed values retain a larger portion of the original cost than those of +// values that have not been accessed for a long time. The C costs are reduced as time +// goes and are restored when accessed. + +template <class Key, class Value> +class GreedyDualSizeCache { + // inner class to define the priority item + class PriorityItem { + public: + PriorityItem(uint64_t actual_priority, uint64_t original_priority, Key key) + : actual_priority(actual_priority), + original_priority(original_priority), + cache_key(key) {} + // this ensure that the items with low priority stays in the beginning of the queue, + // so it can be the one removed by evict operation + bool operator<(const PriorityItem& other) const { + return actual_priority < other.actual_priority; + } + uint64_t actual_priority; + uint64_t original_priority; + Key cache_key; + }; + + public: + struct hasher { + template <typename I> + std::size_t operator()(const I& i) const { + return i.Hash(); + } + }; + // a map from 'key' to a pair of Value and a pointer to the priority value + using map_type = std::unordered_map< + Key, std::pair<ValueCacheObject<Value>, typename std::set<PriorityItem>::iterator>, + hasher>; + + explicit GreedyDualSizeCache(size_t capacity) : inflation_(0), capacity_(capacity) {} + + ~GreedyDualSizeCache() = default; + + size_t size() const { return map_.size(); } + + size_t capacity() const { return capacity_; } + + bool empty() const { return map_.empty(); } + + bool contains(const Key& key) { return map_.find(key) != map_.end(); } + + void insert(const Key& key, const ValueCacheObject<Value>& value) { + typename map_type::iterator i = map_.find(key); + // check if element is not in the cache to add it + if (i == map_.end()) { + // insert item into the cache, but first check if it is full, to evict an item + // if it is necessary + if (size() >= capacity_) { + evict(); + } + + // insert the new item + auto iter = + priority_set_.insert(PriorityItem(value.cost + inflation_, value.cost, key)); + // save on map the value and the priority item iterator position + map_[key] = std::make_pair(value, iter.first); + } + } + + arrow::util::optional<ValueCacheObject<Value>> get(const Key& key) { + // lookup value in the cache + typename map_type::iterator value_for_key = map_.find(key); + if (value_for_key == map_.end()) { + // value not in cache + return arrow::util::nullopt; + } + PriorityItem item = *value_for_key->second.second; + // if the value was found on the cache, update its cost (original + inflation) + if (item.actual_priority != item.original_priority + inflation_) { + priority_set_.erase(value_for_key->second.second); + auto iter = priority_set_.insert(PriorityItem( + item.original_priority + inflation_, item.original_priority, item.cache_key)); + value_for_key->second.second = iter.first; + } + return value_for_key->second.first; + } + + void clear() { + map_.clear(); + priority_set_.clear(); + } + + private: + void evict() { + // evict item from the beginning of the set. This set is ordered from the + // lower priority value to the higher priority value. + typename std::set<PriorityItem>::iterator i = priority_set_.begin(); + // update the inflation cost related to the evicted item + inflation_ = (*i).actual_priority; + map_.erase((*i).cache_key); + priority_set_.erase(i); + } + + map_type map_; + std::set<PriorityItem> priority_set_; + uint64_t inflation_; Review comment: I think its better to handle this case correctly. If it is posssible to overflow just reassign all the entries to original cost and inflation to zero. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
