Object groups may be better than pages

Abstract

I argue against trying to solve the problem of clustering objects into disk pages. Instead, I propose that objects be fetched in groups that may be specific to an application or user, and that can be computed at fetch time. I briefly describe crystals, which serve to record such groups statically. Finally, I speculate that object fetching may be particularly relevant for providing services to very small machines with limited memory, such as personal digital assistants. ContextIn a distributed object-oriented database system, server machines store persistent objects shared by applications running on client machines. When an application invokes an operation on a persistent object, that operation must run on either the server or client. I consider the case where objects are moved or copied to the client for at least the duration of the client transaction. A number of existing object-oriented databases work partially or entirely in this mode of executing operations on the client machine: examples are O2 A number of systems transfer pages from server to client. Transferring pages simplifies storage management, since the units transferred are of fixed size, and it allows the use of hardware page-fault mechanisms. However, this approach requires that one have tremendous faith in mechanisms for clustering objects into pages in a reasonable way. Sophisticated Clustering is a Zero-Sum GameAssume for the moment that we have perfect predictive information about access patterns of applications. There is a certain amount of clustering that we can do that benefits all of the applications, but after this point clustering becomes a zero-sum game: making the clustering better for one application makes it worse for another. This effect is not just an intuition, but was apparent in recent work by Tsangaris and Naughton [12]. If it were possible, we would like to cluster as well as we can for each application. In particular, we should be able to accommodate a new application or a dramatically changed access pattern in a relatively straightforward and inexpensive way.

Disconnected Operation in the Thor Object-Oriented Database System

“…The set of object references captured in this way can be stored and later used in a hoarding query. Such a tool has been proposed for object prefetching [4]; it is similar to the file tracing tool provided by Coda [22] that helps users determine which "hidden files" are used by an application so that they can be included in their hoarding profiles. However, an OODB trace can include not only object identifiers but descriptions of the queries used during the trace run.…”

Disconnected operation in the Thor object-oriented database system

“…The set of object references captured in this way can be stored and later used in a hoarding query. Such a tool has been proposed for object prefetching [4]; it is similar to the file tracing tool provided by Coda [22] that helps users determine which "hidden files" are used by an application so that they can be included in their hoarding profiles. However, an OODB trace can include not only object identifiers but descriptions of the queries used during the trace run.…”