Section: New Results

Transparent resource management

Client-side cloud broker.

Integrating the `pay-as-you-go` pricing model commonly used in IaaS clouds is an important question which profoundly changes the assumptions for job scheduling. From the observation that in most commercial solutions the price of a CPU cycle is identical, be the CPU a fast or slow one, several schedulings may be derived for a same price but with different makespans. Hence, in a context where resources can be started on-demand, scheduling strategies must include a decision process regarding the scaling (number of resources used) of the platform and the types of resources rented over time. In [24] , we have studied the impact of these two factors on classic job scheduling strategies applied to bag-of-tasks workloads. The results show that shorter makespans can be achieved through scaling at no extra cost, while using quicker CPUs largely increases the price of the computations. More importantly, we show the difficulty to predict the outcomes of such decisions, which requires to design new provisioning approaches.