Better Together: Quantifying the Benefits of the Smart Network

"Abstract—Three approaches to load balancing in a distributed computing network are evaluated analytically using statistics and via Monte Carlo simulation: 1) random selection; 2) selection based solely on identifying the server with the lowest response time; 3) selection based on identification of a combination of path and server with the lowest total response time. Analytical and simulation results show that the lowest expected response times occur via joint optimization. The exact improvement depends on the underlying distributions of path response times and server response times. An exemplary case where each path and server is an independent, identically distributed random variable with a continuous uniform distribution on [0,1] is assessed; then the improvement may be approximated by (1/2)+(1/(n+1))-√(π/2n), where n is the number of alternative combinations of path and server. Generally, when P is a random variable representing path response times, the value of a smart network in reducing response time is μ(P)-min(P) as n→∞. Such improvements support a philosophy of “smart” networks vs. “dumb pipes.”..."

As Time Goes By: The Law of Cloud Response Time

"We propose a Law of Cloud Response Time that combines network latency and parallel processing speed-up in a distributed, elastic, cloud computing environment. As the first supercomputing and parallel processing systems came into existence in the 1960s, Gene Amdahl proposed Amdahl's Law: the maximum possible speedup due to parallelization is 1/S, where S is the sequential percentage of the application. Thus, at least for somewhat parallelizable applications, more processors mean less elapsed time, but there is a limit to the gains as no acceleration can occur in the serial portion of the application.

However, today's geographically dispersed cloud environments comprising networked nodes of elastic resources are very different than the local, monolithic, centralized environments of a half century ago, so we propose a new law for interactive transactions over a network with parallelization..." 

Cloud Computing is NP-Complete

"We show that an abstract formulation of resource assignment in a distributed cloud computing environment, which we term the CLOUD COMPUTING demand satisfiability problem, is NP-complete, using transformations from the PARTITION problem and 3-SATISFIABILITY, two of the “core” NP-complete problems. Specifically, let there be a set of customers, each with a given level of demand for resources, and a set of servers, each with a given level of capacity, where each customer may be served by two or more of the servers. The general problem of determining whether there is an assignment of customers to servers such that each customer’s demand may be satisfied by available resources is NP-complete..."

The Market for 'Melons': Quantity Uncertainty and the Market Mechanism

"Using agent-based simulation and analysis of an idealized model of a duopoly with one flat-rate and one usage-based provider, we demonstrate that flat-rate plans are unsustainable in a perfectly competitive market with independent, decentralized decision-making by active, self-selecting, rational utility maximizers engaged in a stochastic, multi-step decision process driving iterative price adjustment. In distinction to Nobel Laureate Dr. George Akerlof’s quality uncertainty in “The Market for ‘Lemons’”, where the seller is advantaged by asymmetric information regarding the quality of the product or service being sold, in what we’ll call “The Market for ‘Melons’” (PDF) it is the buyer that may be advantaged by asymmetric information regarding the ex-ante quantity of planned consumption. Moreover, adverse selection and moral hazard have less to do with quality uncertainty, information asymmetry, or morality, than with rational choice by consumers with dispersed consumption under flat-rate pricing..."