Cloud Pricing: The Spot Market Strikes Back

Abstract

Cloud computing providers must constantly hold many idle compute instances available (e.g., for maintenance or for users with long-term contracts). A natural idea, which should intuitively increase the provider’s profit, is to sell these idle instances on a secondary market, for example, via a preemptible spot market. However, this ignores possible “market cannibalization” effects that may occur in equilibrium as well as the additional costs the provider experiences due to preemptions. To study the viability of offering a spot market, we model the provider’s profit optimization problem by combining queuing theory and game theory to analyze the equilibria of the resulting queuing system. Our main result is an easy-to-check condition under which a provider can simultaneously achieve a profit increase and create a Pareto improvement for the users by offering a spot market (using idle resources) alongside a fixed-price market. Finally, we illustrate our results numerically to demonstrate the effects that the provider’s costs and her strategy have on her profit. This paper was accepted by Gabriel Weintraub, revenue management and market analytics.

A Framework for Allocating Server Time to Spot and On-Demand Services in Cloud Computing

“…Abhishek et al and Dierks et al analyze the performance of a hybrid spot and on-demand market using queuing theory and game theory in continuous time [1,10]. More recently, in [10], the servers are separated into two parts, respectively serving on-demand and spot jobs. Jobs of users have diverse values and sensitivities to delay.…”

“…The authors show that offering a spot market can increase the profit of a CSP. In spite of the technical merits of [1,10], an issue in these works is that in the proposed schemes the idle servers in the on-demand market cannot be sold as spot instances, whereas making use of such idle instances is one of the main attractions of the spot market [9]. Furthermore, inspired by the dynamic pricing of Amazon EC2, there are also many works that apply the auction and mechanism design theory to cloud pricing [12, 30-33, 36, 40-42, 44, 47, 48]; the most relevant ones are [4,16,38].…”

“…We first introduce the framework of Dierks et al in [10,11] where on-demand and spot markets are modeled as two separate queues Q 1 and Q 2 . For Q 1 , the number of servers m o is the minimum severs needed to guarantee that the expected waiting time of jobs is very small; the price of utilizing a sever is p per unit of time.…”

“…A framework covering these aspects is needed in order to offer such services. Existing works [1,10] focus on the case that the spot and on-demand markets are isolated from each other [9]. The pricing of spot instances is studied in [34] where a demand curve is used to describe the relation between the spot price and the number of accepted bids [34].…”

“…However, the curve is unknown in reality [5]. In [1,10,34], the connection of the two markets is ignored; however, the spot market is supposed to utilize the idle servers of on-demand market. Such idleness affects the capacity of spot market to accept bids and the specific schemes for assigning bids to servers.…”