Section: Scientific Foundations

Migration and resilience; Large scale data management

Participant : Yves Denneulin.

Fault tolerance, migration, distributed algorithms.

Most propositions to improve reliability address only a given application or service. This may be due to the fact that until clusters and intranet architectures arose, it was obvious that client and server nodes were independent. This is not the case in parallel scientific computing where a fault on a node can lead to a data loss on thousands of other nodes. The reliability of the system is hence a crucial point. MESCAL's work on this topic is based on the idea that each process in a parallel application will be executed by a group of nodes instead of a single node: when the node in charge of a process fails, another in the same group can replace it in a transparent way for the application.

There are two main problems to be solved in order to achieve this objective. The first one is the ability to migrate processes of a parallel, and thus communicating, application without enforcing modifications. The second one is the ability to maintain a group structure in a completely distributed way. The first one relies on a close interaction with the underlying operating systems and networks, since processes can be migrated in the middle of a communication. This can only be done by knowing how to save and replay later all ongoing communications, independently of the communication pattern. Freezing a process to restore it on another node is also an operation that requires collaboration of the operating system and a good knowledge of its internals. The other main problem (keeping a group structure) belongs to the distributed algorithms domain and is of a much higher level nature.

In order to use large data, it is necessary (but not always sufficient, as seen later) to efficiently store and transfer them to a given site (a set of nodes) where it is going to be used. The first step toward this achievement is the construction of a file system that is an extension of NFS for the grid environment. The second step is an efficient transfer tool that provides throughput close to optimal (i.e. the capacity of the underlying hardware).

Our goal here is to design a distributed file system for clusters that enables one to store data over a set of nodes (instead of a single one). It was designed to permit the usage of a set of disks to optimize memory allocations. It is important for performance and simplicity that this new file system has little overhead for access and updates. From a user point of view, it is used just as a classical NFS. From the server point of view, however, the storage is distributed over several nodes (possibly including the users).