Section: Scientific Foundations

Fondation 1

Scaling to large configurations is one of the major challenges addressed by the distributed system community lately. The basic idea is how to efficiently and transparently use and manage resources of millions of hosts spread over a large network. The problem is complex compared to classical distributed systems where the number of hosts is low (less than a thousand) and the inter-host links are fast and relatively reliable. In such “classical” distributed architectures, it is possible and reasonable to build a single image of the system so as to “easily” control resource allocation.

In large configurations, there is no possibility to establish a global view of the system. The underlying operating system has to make decisions (on resource allocation, scheduling ...) based only on partial and possibly wrongs view of the resources usage.

Scaling introduces the following problems:

• Failure: as the number of hosts increases, the probability of a host failure converges to one. (For instance if we consider a classical host MTBF (Mean Time Between Failure) equals to 13 days, in a middle scale system composed of only 10000 hosts, a failure will occur every 4 minutes.) Compared to classical distributed systems, failures are more common and have to be efficiently processed.

• Asynchronous networks: on the Internet, message delays vary considerably and are unbounded.

• Impossibility of consensus:In such an asynchronous network with failures, consensus cannot be solved deterministically (the famous Fischer-Lynch-Patterson impossibility result of 1985).The system can only approximate, suspecting hosts that are not failed, or failing to suspect hosts that have failed. As a result, no host can form a consistent view of system state.

• Failure models: the classical view of distributed systems considers only crash and omission failures. In the context of large-scale, open networks, the failure model must be generalised to include stronger attacks. For instance, a host can be taken over (“zombie”) and become malicious. Arbitrary faults, so-called Byzantine behaviours, are to be expected and must be tolerated.

• Managing distributed state: In contrast to a local-area network, establishing a global view of a large distributed system system is unfeasible. The operating system must make its decisions, regarding resource allocation or scheduling, based on partial and incomplete views of system state.

Three architectures in relation with the scaling problem have emerged during the last years:

Grid computing:

Grid computing offers a model for solving massive computational problems using large numbers of computers arranged as clusters interconnected by a telecommunications infrastructure as internet or renater.

If the number of involved hosts can be high (several thousands), the global environment is relatively controlled and users of such systems are usually considered safe and only submitted to host crash failures (typically, Byzantine failures are not considered).

Peer-to-peer overlay network:

Generally, a peer-to-peer (or P2P) computer network is any network that does not rely on dedicated servers for communication but, instead, mostly uses direct connections between clients (peers). A pure peer-to-peer network does not have the notion of clients or servers, but only equal peer nodes that simultaneously function as both “clients” and “servers” with respect to the other nodes on the network.

This model of network arrangement differs from the client-server model where communication is usually relayed by the server. In a peer-to-peer network, any node is able to initiate or complete any supported transaction with any other node. Peer nodes may differ in local configuration, processing speed, network bandwidth, and storage capacity.

Different peer-to-peer networks have varying P2P overlays. In such systems, no assumption can be made on the behavior of the host and Byzantine behavior has to be considered.

Cloud computing

Cloud computing offers a conceptually infinite amount of computing, storage and network resources for rent. From a user's perspective, cloud computing has many advantages, including low upfront investment and outsourcing of system administration. From the provider's perspective, cloud computing shares many characteristics with both grid computing (e.g., very large geographical and numeric scale, or service-oriented interfaces) and with P2P computing (e.g., self-administration). It also has some unique characteristics, such as systematic virtualisation of all resources, highly variable load, fast elastic adaptation, and quality-of-service objectives that are negociated with clients via SLAs.

Regal is interested in how to adapt distributed middleware to these large scale configurations. We target Grid and Peer-to-peer configurations. This objective is ambitious and covers a large spectrum. To reduce its spectrum, Regal focuses on fault tolerance, replication management, and dynamic adaptation.

We concentrate on the following research themes:

Data management:

the goal is to be able to deploy and locate effectively data while maintening the required level of consistency between data replicas.

System monitoring and failure detection:

we envisage a service providing the follow-up of distributed information. Here, the first difficulty is the management of a potentially enormous flow of information which leads to the design of dynamic filtering techniques. The second difficulty is the asynchronous aspect of the underlying network which introduces a strong uncertainty on the collected information.

Adaptive replication:

we design parameterizable techniques of replication aiming to tolerate the faults and to reduce information access times. We focus on the runtime adaptation of the replication scheme by (1) automatically adjusting the internal parameters of the strategies and (2) by choosing the replication protocol more adapted to the current context.

The dynamic adaptation of application execution support:

the adaptation is declined here to the level of the execution support (in either of the high level strategies). We thus study the problem of dynamic configuration at runtime of the low support layers.