Skip to Main content Skip to Navigation

Garantir la disponibilité et la gestion de la cohérence dans les systèmes géo-distribués

Abstract : Geo-distributed file systems suffer from high latency and network partitions. Because of this, and to ensure high availability, such systems typically commit updates locally, with no latency, and propagate them in the background. Such optimistic replication faces two major challenges: (i) detecting conflicts between concurrent updates and resolving them in a way meaningful for users, while maintaining system integrity invariants; and (ii) supporting legacy applications that are not prepared to deal with concurrency anomalies. Existing optimistic geo-distributed file systems fall short of addressing the challenges. For instance, Dropbox does not support hard links; Andrew File System fails on some concurrent renaming of directories; and all existing systems use automatic conflict resolution that violates the legacy POSIX semantics. We present our solution to the above problems in the design and implementation of a prototype geo-distributed file system, named Tofu. Its design includes a new session abstraction to support the legacy API, while allowing optimistic updates. Unlike previous approaches, our solution is based on a formal model covering all aspects of a Unix-like file system, including directories, inodes, hard links, etc. It is able to detect all conflicts on those data structures, and resolves them in a way that we believe users will find generally reasonable. Experiments show that Tofu is highly scalable, and incurs linear overhead, improving over existing academic and industrial systems.
Complete list of metadatas

Cited literature [69 references]  Display  Hide  Download
Contributor : Vinh Tao <>
Submitted on : Thursday, December 28, 2017 - 11:11:11 AM
Last modification on : Thursday, March 21, 2019 - 2:32:20 PM


Files produced by the author(s)


  • HAL Id : tel-01673030, version 1


Vinh Tao Thanh. Garantir la disponibilité et la gestion de la cohérence dans les systèmes géo-distribués . Computer Science [cs]. Pierre et Marie Curie, Paris VI; Inria Paris; REGAL; Scality, 2017. English. ⟨tel-01673030⟩



Record views


Files downloads