Function request shipping in a database machine environment

Abstract

In ARBRE, a multiprocessor shared nothing database machine, parts of a single transaction must execute at several processing sites. The code fragments which implement the transaction must somehow come to exist at the necessary sites; they must be linked and loaded; and each must be invoked by a coordinator or subordinate transaction thread. We call this mechanism function request shipping, or FRS. The primary concerns in FRS are about the code that implements the function: its language, representation, transport, and execution. We present three alternative mechanisms for code transport and execution: piggyback, callback, and shared disk. The performance of these mechanisms, with and without code caching, was measured on the ARBRE prototype. The results indicate that the method of choice depends on both the cache hit ratio and the size of the code fragments.