Fast Query Processing Using Cooperative CPU Caching for Index Structures

Index data structures are an important component of database applications. Our measurements show that for a memory resident CSB+-tree (Cache Sensitive B+ tree), the cache miss penalty now accounts for over 50% of the total running time. This is exacerbated by current trends. As a result of Moore’s law, both the CPU speed and the DRAM density have doubled every 18 months. However the cache miss penalty for the CPU has increased with today’s longer cache lines. Clusters provide a cost-effective alternative using commodity equipment. We explore a cache conscious index solution for a cluster. We propose a novel approach for fast query processing that aggregates caches over multiple CPUs. We call this cooperative CPU caching. The idea is to decompose the memory-resident sorted array into partitions so that both a partition and an index over that partition can fit into the L2 cache of one CPU. Then we do parallel query processing, in which we trade off cache misses on a single computer for network communication speeds. We evaluate three alternative strategies to query over a cache-resident sorted array. We compare our method using a cluster of n CPUs with the alternative of n independent CPUs, each processing a separate stream of queries. On a standard cluster, our approach is shown to be approximately 50% faster. By managing the aggregate CPU caches, we achieve super-linear speedup. A cluster of 10 CPUs can be more than 10 times faster than a single CPU. In both experimental and theoretical evaluations, we show that our approach outperforms other approaches.