The LFU algorithm evicts pages that have been least frequently used. If the cache is full, and one more page needs to be added, the page that has been used the least number of times is evicted from the cache.
How can we efficiently implement this algorithm? Each operation of the LRU algorithm can be implemented in worst case O(1) complexity. We can do the same for LFU as well.
For this, we maintain a doubly linked list where each node of the linked list looks like this:
CountNode *prev, *next;
Each node of this linked list points to the root of another doubly linked list which holds the actual pages. All the pages in this 2nd linked list have the same usage count. The node for this linked list looks like this:
PageNode *prev, *next;
The PageNode's parent member points to that CountNode of which it is a child. So there is a direct pointer from every PageNode to the corresponding CountNode it falls under. These extra links are not shown in the diagram to keep it simple.
The CountNode linked list's count member holds the usage count of each element under that node. So each PageNode under a CountNode having a count value of 6 will have been accessed 6 times.
How to implement the various operations?
- Insetion: To insert, we always add the new PageNode to the CountNode list with the count value of 1. This will always be the first node in the linked list pointed to by head. If the first CountNode pointed to by head is greater than 1, we just add a new CountNode with a count value of 1 and then perform the insertion. All this can be done in time O(1).
- Deletion: We will delete nodes, only if the cache is full, and this deletion will always happen from the CountNode list which has the least count. So, we just follow head, and delete a PageNode in such a CountNode. Additionally, if that CountNode happens to become empty because of this deletion, we also delete that CountNode. This too takes time O(1).
- Access: Access involves incrementing the count of a PageNode, i.e. moving the accessed PageNode from it's CountNode list to a CountNode list with a count value that is 1 greater than it's current count value. We maintain pointers to the CountNode in the PageNode. This makes deleting the PageNode(and possibly the CountNode to which it belonged) a constant time operation. After that is done, we can insert a new CountNode that has a count value of 1 more than the CountNode from which the accessed PageNode was removed. If such a CountNode does not exist, we can always create it first. All this can also be done in constant time, making the whole process O(1).
A lot of people have asked me how they can get to a PageNode in the first place. The answer to that question is that we maintain a hash table which hashes an object(page address, integer, etc...) such that accessing it (getting to the PageNode that that object corresponds to) can be accomplished in O(1) time.
This implementation takes advantage of the fact that:
- When we add a new object to the cache, it is always going to start off with a use count of 1, and
- Whenever we access an object, it's use count will go up by exactly 1.
|I haven't found any resource that says that LFU can be implemented in time O(1). If you (reader) find any, please do let me know.|