Using data-flow analysis in MAS for power-aware HPC runs

With a growing concern on the considerable energy consumed by HPC platforms and data centers, research efforts are targeting toward green approaches with higher energy efficiency. Hence, the use of low power processors coming from the mobile market (such as ARM or Intel Atom) gains more and more interest [6]. In parallel, novel software approaches are mandatory to effectively use such cutting-edge technologies and dynamically adapt to the execution being performed. This remains one of the biggest challenges to make HPC applications able to take advantage of Exascale platforms once they will be available. In this context, the authors propose a dynamic and flexible scheme based on a Multi-Agent System (MAS) to handle parallel or distributed execution in an HPC environment. The proposed approach uses a portable representation for the distributed execution E of a parallel program on a fixed input: a bipartite Direct Acyclic Graph (DAG) G = (v, ε) known as a macro DataFlow Graph (DFG). The first class of vertices is associated to the tasks (in the sequential scheduling sense) whereas the second one represents the parameters of the tasks (either inputs or outputs according to the direction of the edge). The total number of tasks Tj in G is denoted by |G| = n. A DFG example is proposed in Figure 1. In the following, we will adopt the notation and assumptions of [7]. In particular, G>(T) denotes the sub-graph induced by all successors of a task T ∈ G and G≥(T) = G>(T) ∪ {T}. Tasks in G and therefore E are computed on a distributed computing platform, typically an HPC system. Modelling an execution by a data-flow graph is part of many parallel programming languages and some efficient execution engines such as Kaapi [3] or Cilk [2] use the graph G to schedule and execute programs on distributed architectures.