Self-Adaptation of Parallel Applications in Heterogeneous and Dynamic Architectures

Abstract

In this paper a mechanism for adaptation of parallel computation is defined for data flow computations in dynamic and heterogeneous environments. Our mechanism is especially useful in massively parallel multithreaded computations as found in cluster or grid computing. By basing the state of executions on a data flow graph, this approch shows extreme flexibility with respect to adaptation ofparallel computation induced by application. This adaptation reflects needsfor changing runtime behavior due to time observable parameters. Specifically, it allows an on-line adaptation ofparallel execution in dynamic heterogeneous systems. We have implemented this mechnism in KAAPI (KernelforAdaptative and Asynchronous Parallel Interface) and experimental results show the overhead induced is small.

Optimised Recovery with a Coordinated Checkpoint/Rollback Protocol for Domain Decomposition Applications

“…This work is still in progress, additional evaluations and experiments at grid scale are planned. The final purpose is to provide a framework that adapts dynamically to available resources [11] using the CCK fault-tolerance protocol.…”

“…For the application aimed in this paper, we use the following approach, called static scheduling [11], to execute the data flow graph. First, a pluggable library like SCOTCH [27] or METIS [28] partitions the data flow graph of one iteration in N data flow subgraphs where N is the wanted processor number to run on.…”

“…Exclusive reservation of computing resources during such a period conflicts with reservation policies aiming at fairness between users on short periods. In this case, fault-tolerance allows to split a large computation and run it during many shorter separated reservations [10,11].…”

Impact of Over-Decomposition on Coordinated Checkpoint/Rollback Protocol

“…In the context of this paper, Kaapi schedules an over-decomposed data flow graph using a static scheduling strategy [15,11]. Once computed for a loop body, the data flow graph scheduling can be re-used through several iterations until the graph or the computing resources change.…”