The Second Workshop On Storage, co-organized by Technicolor and INRIA is an open event bringing together experts in the area of distributed & cloud storage, to discuss various aspects of the problem, expose recent progress in the domain, identify research challenges. The topics addressed will include Cloud & data centers architectures, cloud deployments, codes for distributed storage, data consistency, storage.

The workshop is free, but for logistic reasons, registration is required. We still accept late registration for attending the workshop: complete this form as soon as possible. In case of cancellation, or for any question please send a mail at wos@bretagne-networking.org

The workshop will take place at IRISA (INRIA Rennes), in the amphitheater.


9:00 Welcome

9:30-10:50 Session 1 (short talks):

Backup based on coded data and metadata in the cloud, Thomas Mager (Eurecom) [slides]
Efficient deduplication for cluster-based storage systems, Kostas Kloudas (INRIA) [slides]
Efficient maintenance framework for storage systems, Alex Van Kempen (Technicolor)

11:00-12:20 Session 2:

CamCube – Rethinking the Data Center Cluster, Paolo Costa (Imperial College) [slides]
Rhea: automatic filtering for unstructured cloud storage, Dushyanth Narayanan (Microsoft Research) [slides]

14:00-16:00 Session 3:

Challenges in Digital Services Delivery: The Cloud vs. The Crowd, Christophe Diot (Technicolor)
Bipartite Graph Structures for Efficient Balancing of Heterogeneous Loads, Laurent Massoulié (INRIA) [slides]
Design and Analysis of Networks of Caches, Jim Kurose (University of Massachusetts) [slides]

16:15 – 17:35 Session 4:

Geo replication all the way to the edge, Marc Shapiro (INRIA) [slides]
Cloud storage security, Christian Cachin (IBM) [slides]

17:35 – 18:00 Wrap-up


Abstract of Talks

CamCube – Rethinking the Data Center Cluster, Paolo Costa (Imperial College)

Since the early days of networks, a basic principle has been that endpoints treat the network as a black box. An endpoint injects a packet with a destination address and the network delivers the packet. This principle has served us well, and has enabled the Internet to scale to billions of devices using networks owned by competing companies and running applications developed by different parties. However, this approach might not be optimal for large-scale Internet data centers, such as those run by Amazon, Google, Microsoft and Facebook, in which all the components are controlled by a single entity. In the CamCube project, we have been looking at a different approach to build data centers, borrowing ideas from the fields of high performance parallel computing, distributed systems and networking. We use a direct-connect topology, similar to those used in HPC, and a novel networking stack, which supports a key-based routing functionality. By providing applications with a more fine-grain
ed control on network resources, CamCube enables increasing performance and reducing development complexity and cluster costs. In this talk, I will provide an overview of the CamCube platform and motivate its peculiar design choices. I will also describe the design and the evaluation of a number of services that we implemented on CamCube. These include CamKey, a distributed key-value store that takes advantage of the CamCube platform to increase performance and simplifies failure recovery. It also supports a novel caching mechanism to efficiently handle popular queries.

Rhea: automatic filtering for unstructured cloud storage, Dushyanth Narayanan (Microsoft)

Unstructured storage and data processing are increasingly popular for their simplicity, scalability, and flexibility. For this reason, cloud-based scalable storage platforms such as Amazon’s S3 and Azure Storage offer unstructured rather than structured storage. Data are stored in flat files or blobs on the storage layer, and these files or chunks are read in their entirety by application code running on a physically separate compute infrastructure in the same data center.
Transferring data from storage to compute nodes uses scarce and oversubscribed core data center network bandwidth. If data is accessed across providers or data centers then transferring it over the WAN is even more expensive, both for performance and in terms of charges for egress bandwidth.
This talk describes Rhea, a system to automatically generate and run storage-side data filters for unstructured and semi-structured data through static analysis of the computation using the data. Filters are safe, stateless, side effect free, best effort, and transparent to both storage and compute layers. They achieve a reduction in data transfer of 2x-20,000x. This results in reduced network load, shorter job runtimes, and savings on bandwidth.

Challenges in Digital Services Delivery: The Cloud vs. The Crowd, Christophe Diot (technicolor)

The universal answer to home service delivery these days seems to be “The Cloud”, even though nobody really agrees on what “The Cloud” is. In order to bring some transparency to the Cloud, we identify what are the challenges in digital home services delivery, discuss the strengths and limitations of a pure cloud approach, and finally propose an hybrid solution relying both on data centers and home devices to better serve home users. We discuss the research and technology challenges that have to be solved to deploy this digital service delivery architecture.

Bipartite graph structures for efficient balancing of heterogeneous loads, Laurent Massoulié (Technicolor)

This presentation considers large scale distributed content service platforms, such as peer-to-peer video-on-demand systems. Such systems feature two basic resources, namely storage and bandwidth. Their efficiency critically depends on two factors: (i) content replication within servers, and (ii) how incoming service requests are matched to servers holding requested content. To inform the corresponding design choices, we make the following contributions. We first show that, for underloaded systems, so-called proportional content placement with a simple greedy strategy for matching requests to servers ensures full system efficiency provided storage size grows logarithmically with the system size. However, for constant storage size, this strategy undergoes a phase transition with severe loss of efficiency as system load approaches criticality.
To better understand the role of the matching strategy in this performance degradation, we characterize the asymptotic system efficiency under an optimal matching policy. Our analysis shows that -in contrast to greedy matching- optimal matching incurs an inefficiency that is exponentially small in the server storage size, even at critical system loads. It further allows a characterization of content replication policies that minimize the inefficiency. These optimal policies, which differ markedly from proportional placement, have a simple structure which makes them implementable in practice.
On the methodological side, our analysis of matching performance uses the theory of local weak limits of random graphs, and highlights a novel characterization ofmatching numbers in bipartite graphs, which may both be of independent interest.

Design and Analysis of Networks of Caches, Jim Kurose (University of Massachusetts)

Today’s Internet architecture, nearly 40 years old now, is grounded in a model of host-to-host communication. More recently, a number of researchers have begun to focus on Content Networking – a model in which host-to-content (rather than host-to-host) interaction is the norm. Here, content distribution and retrieval, rather than host-to-host packet delivery, is the core function supported in each and every network node. A central component of proposals for such content delivery is the routing of content to requestors through a large-scale interconnected network of caches.In this talk we focus on this cache network. We begin with a quick overview of Content Networking. We then describe Breadcrumbs – a simple content caching, location, and routing system that uses a small amount of information regarding cache history/routing in a simple, best-effort approach towards caching. In the second part of this talk we consider the broad challenge of analyzing networks of interconnected caches.We describe an iterative fixed-point algorithm for approximating cache network performance, evaluate the accuracy of the approximation, and identify the sources of approximation error. We also consider the steady state behavior of cache networks. We demonstrate that certain cache networks are non-ergodic in that their steady-state characterization depends on the initial state of the system. We describe sufficient conditions (based on topology, admission control, and cache replacement policy) for ergodicity and ergodicity equivalence classes among policies. Last, we describe current work on developing a network calculus for cache network flows. Joint work with Elisha Rosensweig, Daniel Menasche, Don Towsley

Geo replication all the way to the edge, Marc Shapiro (INRIA)

Recent developments, such as Ajax and HTML~5, are moving the execution of distributed applications towards client machines. Current data management solutions for cloud infrastructures replicate data among several geographically distributed data centres but lack support for managing data maintained by clients. This paper presents SwiftCloud, a storage infrastructure for cloud environments that covers this gap. SwiftCloud addresses two main issues: maintaining replicas consistent and maintaining client replicas up-to-date. SwiftCloud pushes the scalability and concurrency envelope, ensuring strong eventual consistency using the well-founded approach of Conflict-Free Replicated Data Types (CRDTs). CRDTs provide higher-level object semantics, such as sets, maps, graphs and sequences, support unsynchronised concurrent updates, while provably ensuring consistency, and eschewing rollbacks. Client-side replicas are kept up to date by notifications, allowing client transactions to execute entirely locally, both for queries and for updates. To reconcile performance and reliability, SwiftCloud supports a unified state- and operation-based object representation. We show how we can build significant applications by composing objects from predefined CRDT types using emph{asynchronous transactions}. The programming model is simpler and client-perceived latency is much smaller than with competing approaches.

Cloud storage security, Christian Cachin (IBM)

Cloud storage implies that a data owner loses control over its data in several ways. After the owner outsources its data to the cloud, the owner is bound to trust the cloud provider for confidentiality, privacy, integrity, and availability of its data. Cryptographic mechanisms can reduce such trust by allowing the user to protect its data. This presentation discusses principles of cloud data security and reviews novel mechanisms for securing data stored in the cloud. Such techniques are the focus of current research projects at IBM Research – Zurich and they range from encryption and cryptographic data authentication to protocols for resilient storage and secure deletion.

Workshop Committee:

  • Ernst Biersack (Eurecom)
  • Frédéric Giroire (CNRS)
  • Anne-Marie Kermarrec (INRIA)
  • Erwan Le Merrer (Technicolor)
  • Fabio Picconi (Technicolor)
  • Gilles Straub (Technicolor)

For any question, please mail us at: wos@bretagne-networking.org