We introduce a computation model for streaming and external memory data that distinguishes between sequentially reading (streaming) data from external memory (through main memory) and randomly accessing external memory data at specific memory locations. It is well-known that the latter is much more expensive in practice. We explain how a number of lower bound results are obtained in this model and how they can be applied for proving lower bounds for XML query processing. In the scenario with a single external memory device, strong lower bounds can be obtained by using techniques from communication complexity. To prove tight lower bounds for a setting where two or more external memory devices are available, we simulate computations by non-uniform models called list-machines, for which we can establish lower bounds by combinatorial means.
This is joint work with Martin Grohe, Andre Hernich, and Christoph Koch.