Cloudonomics: The Business Value of Cloud Computing, + Website

Benefits of Consolidation

The statistics of scale, as we saw in Chapter 15, can drive a relative penalty cost effect when m independent workloads or customer demands are aggregated. In other words, as the volume of work grows in proportion to m, the expected absolute penalty only grows as , thus the relative penalty is which is . To the extent that only “local” workloads are aggregated, dispersion causes problems due to “breakage.” There is still a smoothing effect; demand just won’t get as flat as with greater aggregation. Aggregating fewer workloads from within smaller regions, under an assumption that customers can’t be serviced out of more distant nodes, means that the effective m is smaller and thus the relative penalty cost is larger.

Suppose we have m customers, each with, say, normally distributed, independent demands, each with standard deviation σ. Let us partition these customers and thus their demands evenly among n nodes, so that each node has m/n customers. Since the sum ...