Bid selection for deadline constrained jobs over spot VMs in computational cloud

Abstract

Spot instance is an attractive pricing option to the cloud service users. But to use the spot instances, the cloud service users (CSU) have tos select an appropriate bid. The total cost of using spot instances and time to complete the user’s job are both determined by the bid chosen. Thus, selection of appropriate bid for spot instances is a critical problem for the CSUs because of the unpredictable nature of spot prices. The problem is more challenging when there is a deadline constraint. The Dynamic Bidding Algorithm (DBA) was considered to be an effective approach in this field. In this paper, we show that by combining a greedy approach for bidding with effective use of checkpointing, the total cost of using spot instances can be reduced further without violating the deadline constraint. To measure the cost of our solutions and those given by DBA with the optimum we have formulated an IPP for the given problem. In formulating the IPP it is assumed that spot prices at all future times are known. Experimental results based on some actual spot and on-demand price data show that the cost of our solution is comparable to those given by IPP. It is also observed that the cost decreases as the deadline to complete the job is extended.