2Q: A Low Overhead High Performance Buffer Management Replacement Algorithm

In a path-breaking paper last year Pat and Betty O'Neil and Gerhard Weikum proposed a self-tuning improvement to the Least Recently Used (LRU) buuer management algorithmm15]. Their improvement is called LRU/k and advocates giving priority to buuer pages based on the kth most recent access. (The standard LRU algorithm is denoted LRU/1 according to this terminology.) If P1's kth most recent access is more more recent than P2's, then P1 will be replaced after P2. Intuitively, LRU/k for k > 1 is a good strategy , because it gives low priority to pages that have been scanned or to pages that belong to a big randomly accessed le (e.g., the account le in TPC/A). They found that LRU/2 achieves most of the advantage of their method. The one problem of LRU/2 is the processor overhead to implement it. In contrast to LRU, each page access requires log N work to manipulate a priority queue where N is the number of pages in the buuer. Question: is there low overhead way (constant overhead per access as in LRU) to achieve similar page 0 replacement performance to LRU/2? Answer: Yes. Our \Two Queue" algorithm (hereafter 2Q) has constant time overhead, performs as well as LRU/2, and requires no tuning. These results hold for real (DB2 commercial, Swiss bank) traces as well as simulated ones. Based on these experiments, we estimate that 2Q will provide a few percent improvement over LRU without increasing the overhead by more than a constant additive factor. 1 Background Fetching data from disk requires at least a factor of 1000 more time than fetching data from a RAM buuer. For this reason, good use of the buuer can signii-cantly improve the throughput and response time of any data-intensive system. Until the early 80's, the least recently used buuer replacement algorithm (replace the page that was least recently accessed or used) was the algorithm of choice in nearly all cases. Indeed, the theoretical community blessed it by showing that LRU never replaces more than a factor B as many elements as an optimal clair-voyant algorithm (where B is the size of the buuer) 19]. 1 Factors this large can heavily innuence the behavior of a database system, however. Furthermore , database systems usually have access patterns in which LRU performs poorly, as noted by Stone-braker 21], Sacco and Schkolnick 18], and Chou and Dewitt 5]. As a result, there …