2. RELATED WORK
In Big-Data-scale workloads, building a cluster with enough DRAM capacity to accommodate
the entire dataset can be desirable but expensive. An example of such
a system is RAMCloud, which is a DRAM-based storage for large-scale data center
applications [Ousterhout et al. 2010; Rumble et al. 2014]. RAMCloud provides more
than 64TB of DRAM storage distributed across over 1,000 servers networked over
high-speed interconnect. Although RAMCloud provides 100 to 1,000 times better performance
than disk-based systems of similar scale, its high energy consumption and
high price per GB limit its widespread use except for extremely performance- and
latency-sensitive workloads. Furthermore, the overhead of a widely distributed processing
platform becomes high quickly, making it difficult to make full use of the total
computational power of the cluster [McSherry et al. 2015]. Even with a scalable processing
platform, the overhead may be so high that running a less scalable software on
fewer machines might sometimes be a faster solution.