Cooperative Caching in Append-only Databases with Hot Spots

We measure the performance of several cooperative caching policies for a database with hot spots. The workload consists of queries and append-only update transactions, and is modeled after a financial database of stock (historical) trading information. We show that cooperative caching is effective for this application. We show that selecting the correct set of peer servers when servicing a cache miss is crucial to achieving high performance, and we demonstrate a greedy algorithm that performs close to optimal for this workload. We also evaluate several cache replacement policies and show that a 2 nd -chance algorithm performs best. In a 2 nd -chance algorithm, replaced pages are transferred to a peer server rather than being discarded. When a page is selected for replacement a 2 nd time, the page is discarded. Our results can be applied in the design of "proxy" servers for databases or web servers where a layer of proxy servers are used to scale the system performance. 1. Introdu..