Skip to Main content Skip to Navigation

Calcul Haute-Performance et Mécanique Quantique : analyse des ordonnancements en temps et en mémoire

Abstract : This work shows the importance of scheduling for the programming of high performance numerical applications of quantum physics and chemistry. We focus on two methods for the Schrödinger equation : the Quantum Box (QB) and the order two Moeller-Plesset (MP2) perturbation algorithm. Both require very large amounts of computing time and memory. The traditional programming (message passing and/or multithreading) of parallel architectures (distributed or SMP) is illustrated through the performances obtained with the Linpack benchmark on the I-Cluster machine of INRIA. The lack of portability of the high-performance code obtained shows the need for an environment for parallel programming that allows the disconnecting of algorithm coding from its scheduling on the target machine. We then introduce Athapascan, that relies on the analysis of the data flow to dynamically compute some schedules of proven efficiency. For the execution time on some machine models, both theory and experiments show that Athapascan provides some schedules that guarantee good performance for algorithms adapted to QB, of iterative nature (Lanczos algorithm). Another aim is to bound the memory required by parallel executions in numerical computing ; this is especially true for MP2. We propose to annotate the Data Flow Graph (DFG) to take the memory into account and allow scheduling that is time and memory efficient. For MP2, whose DFG is statically known, a memory and time efficient scheduling is given.
Document type :
Complete list of metadatas
Contributor : Thèses Imag <>
Submitted on : Monday, February 16, 2004 - 2:28:33 PM
Last modification on : Friday, November 6, 2020 - 4:39:01 AM
Long-term archiving on: : Friday, April 2, 2010 - 8:19:41 PM


  • HAL Id : tel-00004684, version 1



Nicolas Maillard. Calcul Haute-Performance et Mécanique Quantique : analyse des ordonnancements en temps et en mémoire. Modélisation et simulation. Université Joseph-Fourier - Grenoble I, 2001. Français. ⟨tel-00004684⟩



Record views


Files downloads