An online mechanism for dynamic instance allocation in reserved instance marketplace

As one of the pricing model offered by Amazon, reserved instance enables users to reserve capacities for their EC2 instances and lowers their average instance cost. To attract more users to adopt the reserved instance, Amazon has provided a platform named Reserved Instance Marketplace to give users the flexibility to sell the remainder of their reserved instances as their needs change. However, the present trading mechanism inside the Reserved Instance Marketplace requires the reserved instance sellers to sell their reserved capacities at month level and set the upfront fee by themselves, which is not flexible enough and hard for an inexperienced seller to specify a suitable upfront fee. To address this problem, this paper proposes an online mechanism for the Reserved Instance Marketplace. Our online mechanism tries to maximize the sellers' revenue by dynamically allocating the reserved capacities among various buyers without the need of specifying upfront fee in advance. The competitive ratio of the online allocation algorithm inside our mechanism is proved to be within a small constant factor of optimal competitive ratio in theory. To evaluate the performance of our online mechanism, we conduct simulations on synthetic data and real data trace from one of Google clusters. The simulation results show that our online mechanism can achieve at least 55% of the offline optimal algorithm in most cases.