08.02.15 - 11.02.15, Seminar 15072

Distributed Cloud Computing

Diese Seminarbeschreibung wurde vor dem Seminar auf unseren Webseiten veröffentlicht und bei der Einladung zum Seminar verwendet.

Motivation

The applications and usage patterns of the Internet have changed significantly over the years. One of the most striking properties of the Internet today regards the enormous amount of traffic it carries, imposing high bandwidth requirements and costs. Also latency is becoming increasingly critical. Motivated by these challenges, the distributed cloud model envisions new architectures exploiting spatially distributed resources, to keep traffic more local and reduce latency, offering innovative new services. For example, today, various resources are already available in the geographically distributed facilities of telecoms. These "resource micro clouds" could be used and interconnected to offer new services in the network core. Distributed cloud also encompasses the federated cloud, where data centers managed by different organizations federate to allow users to utilize any of the data centers.

Goals

This interdisciplinary seminar is an opportunity for academic and industrial scientists working in the different fields of networking, cloud computing, and distributed systems, to meet and exchange their vision and expertise on how to plan and manage future research on distributed cloud. Accordingly, we plan to structure our seminar in three parts, cloud, networking, and distributed systems, and complement the participant presentations with panels that give room to exchange and challenge ideas. We expect this seminar to formulate a concrete research agenda in the design issues and standards of distributed clouds.

Research Topics

The two most distinctive features of the distributed cloud are the heterogeneous nature of resources and providers, and the relatively limited networking (relative to clouds based in a single data center). These two features lead to a number of distinct research questions, including:

  1. Federated vs integrated vs autonomous? What are the tradeoffs of these distributed cloud models?
  2. Enablers for a more dynamic distributed cloud management
  3. Monitoring, allocation, and control of virtual machines in the presence of network failures and when latencies are in the tens to hundreds of milliseconds
  4. Programming models suitable for the wide area - notably, models which are optimized for low bandwidth and high latency, where sending programs to data (rather than data to programs) is of paramount importance
  5. Storage systems suitable for the wide area, with emphasis on caching, robustness, and location awareness
  6. Concurrency issues and scalability of the control plane
  7. Incremental deployment strategies
  8. Service differentiation a la Google B4
  9. Distributed operating system vs distributed cloud management (e.g. distributed OpenStack) approaches - advantages and disadvantages to both
  10. Distributed agreement algorithms like RAFT and Paxos, scalability and applicability to specific use cases
  11. Better support for network virtualization going out of the data center and between data centers
  12. Drill down on these use cases
    a. Downstream large data streaming, e.g. CDN
    b. Upstream large data summarization, e.g. voice recognition
    c. Low latency interactive applications, e.g. virtual reality (7-10 ms), industrial control
    d. Core network offload - reducing amount of traffic in carrier’s core network by turning it around in edge distributed clouds
  13. Economics of distributed cloud deployment: cheaper or more expensive than centralized cloud? Cost of networking resources for distributed cloud vs centralized cloud?
  14. Deployment models for cloud: hyper-centralized (e.g. one data center per continent), centralized (two or three data centers per continent), distributed (one data center per metro area), hyper-distributed (a data center on your street corner). In what cases is a specific deployment model favored?