In this section, we describe how architectural decisions made at
the system-level affect the relative performance of the two classes of
data analysis systems. Since installation and configuration parameters can have a significant difference in the ultimate performance
of the system, we begin with a discussion of the relative ease with
which these parameters are set. Afterwards, we discuss some lower
level implementation details. While some of these details affect
performance in fundamental ways (e.g., the fact that MR does not
transform data on loading precludes various I/O optimizations and
necessitates runtime parsing which increases CPU costs), others are
more implementation specific (e.g., the high start-up cost of MR).