Scheduling/Data Management Heuristics

Frédéric Desprez (1) , Sylvain Gault (1) , Frédéric Suter (2, 1)

1 AVALON - Algorithms and Software Architectures for Distributed and HPC Platforms
2 CC-IN2P3 - Centre de Calcul de l'IN2P3

Data volume produced by scientific applications increase at a high speed. Some are expected to produce several petabyte per year. In order to process this amount of data, the computing power of several hundreds or thousands of machines have to be used at the same time. Regarding this, one of the biggest challenge is: how to program these machines in order to make them to collaborate for the same computation? One answer brought by Google is the MapReduce paradigm. MapReduce has the advantage of being quite simple to program for the user and handle on its own the repetitive or complex tasks like the data transfers between nodes, task scheduling or handling node failure. These automatic tasks have to be handled in an optimized way in order to make the framework fast and scalable. This report presents our first studies towards an efficient scheduling of MapReduce operations. More specifically, we focused on the scheduling of the data transfers together with the tasks. We present here an interesting work around this topic and our algorithm which improves their results.

Domaines: Calcul parallèle, distribué et partagé [cs.DC]

Frédéric Desprez, Sylvain Gault, Frédéric Suter. Scheduling/Data Management Heuristics. 2012. ⟨hal-00759546⟩