Monday 19th June – INRIA : Guillaume Pallez (INRIA)

Title: Stochastic Profiles of upcoming HPC Applications and their impact on Scheduling strategies

Abstract:

New applications with different profiles are coming to HPC infrastructures. We aim at better understanding the features and needs of these applications in order to be able to run them efficiently on HPC platforms.

In this talk I will present how we studied an emerging application, Spatially Localized Atlas Network Tiles (SLANT, originating from the neuroscience community) to understand its behavior. Based on these observations, we derive a generic, yet simple, application model (namely, a linear sequence of stochastic jobs). With respect to the Inria-LJLL seminar, I am interested to understand how these application models could apply to the applied math community.

In a second step, we show how one can use this generic model in a scheduling framework. Specifically we consider the problem of making reservations for an execution on an HPC platform based on the application expected resource requirements. We derive solutions using the model provided by the first step of this work. We experimentally show the robustness of the model, even with very few data points or using another application, to generate the model, and provide performance gains with regards to standard and more recent approaches used in the neuroscience community.

Comments are closed.