Cache conscious clustering C3

Abstract

Despite over 10 years of research into OODBMS design, performance remains as one of the major problems. I/O reduction has proven to be one of the most effective ways of enhancing performance. Two techniques of improving I/O performance of OODBMS are clustering and buffer replacement. Clustering places objects that are likely to be accessed together onto the same disk page and thus reduces I/O. Buffer replacement involves the selection of a page to be evicted, when the buffer is full. The page evictedid eally shouldb e the page neededfurthest in the future. Selection of the correct page for eviction results in a reduction in the total I/O generatedb y the system. These two techniques effect the likelihoodof a requestedob ject being memory resident in an interdependent way. This fact has been acknowledged in the existing literature. However despite this acknowledgement no existing work investigates this interdependency. This paper makes the first investigation into this interdependency by exploring the effects of ten different buffer replacement algorithms on the performance of two different clustering algorithms. We developed a new family of clustering algorithms that incorporate cache behavior when performing clustering. We term this new family of clustering algorithms, cache conscious clustering (C3). A particular C3 algorithm (GPC3) was testedagainst a well known andhighly competitive clustering algorithm GGP, andthe results show GPC3 outperforms GGP by upto 40% for popular buffer replacement algorithms such as LRU andCLOCK. These results show for the first time clustering shouldb e approachedfrom a cache conscious point of view.