Optimization of compute infrastructure costs is an actively studied area. Multiple proposals advocate the use
of low-power, energy efficient cores for datacenter workloads. Reddi et al. [26] quantified the cost of efficiency
when running a search application on heavyweight vs. lowpower cores. They show that low-power cores like the Intel
Atom show improvements in power efficiency when com pared to a server class CPU like the Intel Xeon. This efficiency however comes at the expense of quality-of-service
guarantees due to the increase in latency and variability
for search queries. They suggest over provisioning and
under-utilization of Atom cores and provide evidence that
even with over-provisioning, Atom based platforms incur
only 0.4x the platform-level cost and power efficiency of
the Xeon. Our scheme follows a similar chain of thought
where a large number of Atom CPUs are used in the cluster. They also propose building multi-socket systems using Atom CPUs to amortize the platform level costs. This
also is in line with our proposal since our scheme amortizes
platform costs, but does so more efficiently by building independent compute nodes instead of building a multi-socket
SMP system. Multi-socket SMP systems have higher capital costs compared to our proposal. It is an open research
problem to investigate if multi-socket SMP systems deliver
higher performance for scale-out architectures, or if single
socket systems are better.