Outperforming LRU via Competitive Analysis on Parametrized Inputs for Paging

Abstract

Competitive analysis was often criticized because of its too pessimistic guarantees which do not reflect the behavior of paging algorithms in practice. For instance, many deterministic paging algorithms achieve the op

Engineering Efficient Paging Algorithms

“…In ONOPT, we singled out RDM, which combines two priority policies, one based on recency and the other on the timeframe that pages spent in support. RDM achieves good results, outperforming LRU on many real-world traces and cache sizes [Moruz and Negoescu 2012].…”

“…Although in general C OPT is not known since it depends also on the future request sequence τ , we are provided with partial information (from σ ) about the structure of C OPT ; for example, it contains for sure the most recently requested page, and pages not requested in σ are not in C OPT . We say that immediately after processing σ, a set C of k pages is a valid configuration if and only if there exists a future request sequence τ such that LFD's cache content equals C. A precise mathematical characterization of all possible valid configurations was given by Koutsoupias and Papadimitriou [2000], and an equivalent variant of this characterization is used by the OnOPT algorithm class [Moruz and Negoescu 2012]. It consists of a partition L = (L 0 | .…”

“…Based on the layer partition in Koutsoupias and Papadimitriou [2000], we proposed in Moruz and Negoescu [2012] a measure quantifying the "evilness" of the adversary that we denoted attack rate. For inputs having attack rate r, we introduced a class of r-competitive algorithms, denoted ONOPT, and we showed that these algorithms achieve a small fault rate on many practical inputs (LRU is also in ONOPT).…”

“…Algorithms outperforming LRU in terms of cache misses exist but are in general more complex and thus not automatically better, since their increased runtime might annihilate the gains in cache misses. In this article, we focus on efficient implementations for the ONOPT class described in Moruz and Negoescu [2012], particularly on an algorithm in this class, denoted RDM, that was shown to typically incur fewer misses than LRU. We provide experimental evidence on a wide range of cache traces showing that our implementation of RDM is competitive to LRU with respect to runtime.…”

Improved Space Bounds for Strongly Competitive Randomized Paging Algorithms

“…This gap is significantly smaller for randomized algorithms, since the best possible competitive ratio is H k . Although using only the quality guarantees provided by competitive analysis is a naive way to distinguish good paging algorithms from bad ones, we have shown in [14] that ideas from competitive analysis for randomized algorithms can be successfully employed to design algorithms with good performance on real-world inputs. That is because an optimal randomized algorithm can be viewed as a collection of reasonable deterministic algorithms, and the algorithm designer can simply look for suitable algorithms in this collection.…”