ALTA: Asynchronous Loss Tolerant Algorithms for Grid Computing
Abstract
This paper describes an environment dedicated to the building of efficient scientific applications for the Grid on top of unreliable communication networks. Nowadays, scientific computing appli-cations are usually built on top of reliable communication proto-cols (such as TCP). Nevertheless, the additional cost introduced by the reliability layer is not negligible in wide area network-based grid environments. On the other hand, data loss in communications may have a dramatic impact over the performance – if not over the correctness – of classical parallel algorithms. However, a particular class of parallel iterative algorithms hap-pens to be tolerant to such losses. This is the class of asynchronous iterative algorithms, which are commonly used in large scientific applications. They are particularly prone to a good communica-tion/computation overlap since processors are no more synchro-nized. In this study, we aim at proposing a new architecture suit-able for the development of asynchronous iterative algorithms tolerant to message losses.
Format | Short paper |
---|---|
Origin | Files produced by the author(s) |