In an OR-DBMS, multi-megabyte attributes generally store only metadata with
in the tuple; the data itself is in external objects. The objects and the t
uples generally start on the same node in a shared-nothing system, but tupl
es move around without their data during query processing. In this paper, w
e study how to execute a method on an attribute when the tuple is on one no
de, and the data is on another. We study four techniques, based on whether
to ship the method or the data and whether to ship one-by-one or in larger
units. We present a series of performance studies, which establish that the
effect of the various solutions on the disk utilization is far more signif
icant than their effect on the network load. (C) 2000 Elsevier Science Ltd.
All rights reserved.