Datacenter scheduling of Spot virtual machines
Spot instances have been recently introduced by cloud service providers to monetize the residual capacities in their datacenters. These instances are sold off as preemptible VMs with no SLAs at a discounted rate. Any scheduler that supports both on-demand and spot instances has to balance two competing goals. The scheduler must continue to serve on-demand instances with little delay, while also ensuring that spot instances should be able to run for some minimum amount of time. An effective scheduling strategy would allow for much greater resource utilization and revenue without degrading on-demand scheduling quality.
Much of the prior work on spot instances has been focused on investigating pricing mechanisms. Some of this work has tried to analyze commercial cloud service providers’ pricing strategies whereas other work has proposed mechanisms for greater revenue. There has also been some prior work on workload allocation between on-demand and spot instances from the user’s perspective and also on providing probabilistic SLAs on spot instance lifetimes. The problem of co-scheduling on-demand and spot instances from the cloud service provider perspective has not been satisfactorily addressed; prior attempts to solve the problem have oversimplified it or are too shallow to be useful.
This work is a continuation of our project tackling scheduling with network guarantees. We expect the constraint solving approach we’ve applied to that problem to allow us to explore several promising directions to this research question; success will, as expected, depend on the accuracy of the formal model.
Systopia lab is supported by a number of government and industrial sources, including Cisco Systems, the Communications Security Establishment Canada, Intel Research, the National Sciences and Engineering Research Council of Canada (NSERC), Network Appliance, Office of the Privacy Commissioner of Canada, and the National Science Foundation (NSF).