Public cloud providers are exploring the use of resource overbooking for improving their profitability. As overbooking in the airline industry helps companies sell more tickets and thereby ensure fuller flights, overbooking in the public cloud helps providers create and sell more virtual machines (instances). While such instances are cheaper, they are more difficult to use because their resource capacities vary over time. The goal of this project is to develop an understanding of such burstable instances - so called due to the sporadic or 'bursty' nature of their capacities - and help tenants run their workloads cost-effectively. The project will engage in outreach to industry, particularly to continue active collaborations with IBM, Google, and Amazon. Efforts will be made to convey enhancements to provider-side consolidation techniques to these industrial partners via student internships, talks, and complementary proposals for research/teaching credits. The project will involve under-represented minorities in computer science and engineering, particularly women graduate students. Finally, the work-product of the project will be publicly disseminated on the web, together with developed code, to ease experimental reproduction and validation by other parties.
The project will develop fundamentally novel methodologies for inferring undisclosed parameters governing resource regulators (e.g., variants of token buckets) for burstable instances. Integrating such situational awareness of effective capacity and employing it in application programming platforms (memcached, Spark, and Hadoop) would pose significant challenges since a plurality of resource management strategies in these platforms will now need to co-exist. A final set of research contributions will address novel provider-side concerns related to enriching future burstable instances with more diversity and flexibility for tenants taking into account their more careful usage of these instances. A website for the project will be created at the following URL: http://www.cse.psu.edu/~bhuvan/burstables.html. This website will contain all source code (tenant workload enhancements, provider-side enhancements in Apache Mesos or kubernetes, scripts for running and interpreting experiments, etc.), data, and documents (technical reports, code descriptions, papers at conferences and journals) resulting from the project. The contents will remain available for researchers well beyond the duration of the project.
|Effective start/end date
|9/1/17 → 8/31/20
- National Science Foundation: $500,000.00