Colloquium Details
Optimizing Cloud Resource Allocation and Reconfiguration
| Author: | Lei Jiao Nokia Bell Labs, Ireland |
|---|---|
| Date: | June 07, 2016 |
| Time: | 15:30 |
| Location: | 220 Deschutes |
Abstract
Resource allocation and reconfiguration is fundamental in cloud computing, and becomes more challenging as cloud services and architectures continue to evolve with complexities. In the service aspect, online social networks are popular, and allocating storage for users' data over multiple clouds needs to address the interconnection of users, the master-slave data replication, the conflicting requirements of different objectives, as well as the diversity of multi-cloud data access policies. In the architecture aspect, the emerging edge-core tiered structure of distributed clouds complicates the allocation of resources such as virtual machines and servers, due to the reconfiguration cost associated to changing resource allocation decisions over time, the often unpredictable service demands and resource prices, as well as the heterogeneity of resources.
In this talk, I will introduce our work on exploiting optimization to solve these two problems, while addressing all the aforementioned challenges. The first problem is a large-scale combinatorial optimization problem regarding social network data placement over clouds, potentially with multiple system objectives such as carbon footprint, latency, and inter-cloud traffic. We propose to decompose it into two sub-problems of placing master data replicas and slave data replicas respectively, where we leverage the graph cuts technique to solve the master placement problem, use a greedy method to place the slaves, and solve the two sub-problems alternately in multiple rounds. The second problem is an online optimization problem regarding cloud and network resource allocation over time to process dynamic workloads in a tiered infrastructure, where a resource allocation decision needs to be made on the fly without knowing the entire inputs over the time horizon. With no knowledge of future inputs, we leverage the regularization technique to design an online algorithm that decomposes the original problem by constructing a series of sub-problems solvable at each corresponding time slot, with provable worst-case performance guarantees. Experiments with real-world data traces confirm that our approaches have substantial advantages over the state of the arts.
Biography
Lei Jiao is a post-doctoral researcher at Nokia Bell Labs in Dublin, Ireland. His research interests are broadly in theories and algorithms (e.g., optimization, control, game theory), with a focus on modeling, analyzing, and solving real-world problems in distributed and networked systems. His research appeared in conferences such as IEEE INFOCOM, ICNP, IPDPS, and in journals such as IEEE/ACM Transactions on Networking. He was also a recipient of the Best Paper Award in IEEE LANMAN 2013. He received the Ph.D. degree in computer science from University of Göttingen in Göttingen, Germany in 2014. Prior to his Ph.D. study, he was a researcher at IBM Research in Beijing, China.