As computer architects we know that there are extenuat-
ing circumstances to the latter three infractions: (i) The geo-
metric mean has its pros, cons [4], and detractors [3, 11]; (ii)
Existing simulation infrastructure, consistency with previ-
ous research, and slow adoption, are understandable reasons
for using the older benchmark; (iii) Simulating only part of
a benchmark is almost unavoidable given the very long sim-
ulation time required for the tens of billions of instructions
required to execute the reference datasets of CPU2000. This
problem has been dealt with in several papers [5, 6, 8] and
several solutions that include reduced datasets or selective
sampling have been suggested;
At this point we will distinguish between the use of
the integer suite (CINT2000) and the floating-point suite
(CFP2000). The composing benchmarks are generally dif-
ferent: CFP2000 is dominated by easier to parallelize loops
as opposed to CINT2000 which is plagued by control and
data dependencies [4]. Most researchers tend to focus on
extracting additional performance from CINT2000. There-
fore, the major emphasis of this paper is on the partial use
of the CINT2000 suite, the extent of the phenomena, why
it can cause misleading results, and how it can be avoided
(section 5 will briefly explore the use of CFP2000 and show
where we think it should be used).