reduced[1], [2], [3], [4]. However, few of existing application
scheduling approach has considered optimizing the power consumption of the network infrastructure. In a typical commodity
large-scale systems, the network infrastructure is built to
provide high bisection bandwidth for applications that span the
entire system
. However, sharing by diverse applications and
each utilizing a part of the system, many of these network links
are unused. A network usage record from the Fusion cluster
at the Argonne National Laboratory of the U.S., spreading
over around 18 months from November 2009 to May 2011,
testifies such phenomenon (Fig. 1). As shown by this record,
the network utilization can vary significantly depending on
the forwarding policy (in this case ranging from an average to
54.7% to an average of 90.3%). This instance implies that, by
considering the characteristics of the network infrastructure
in application scheduling, more energy conservation can be
achieved