Abstract Details
Abstracts
Author: Guangye Chen
Requested Type: Poster
Submitted: 2019-02-22 11:38:32
Co-authors: Luis Chacon
Contact Info:
Los Alamos National Laboratory
PO box 1663
Los Alamos, NM 87545
USA
Abstract Text:
With ever-increasing computing power and memory capacity, particle check-pointing in particle-in-cell simulations to deal with computer hard-faults is stressing I/O subsystems, and becoming prohibitive as a recovery strategy. Nevertheless, future exascale computers are expected to be significantly more vulnerable to hard faults than current HPC systems, making a viable recovery strategy absolutely essential. To address this conundrum, we consider compression of the particle distribution function (PDF) by unsupervised machine learning [1]. Specifically, we have developed a check-pointing strategy that approximates the PDF with a Gaussian mixture [2]. The Gaussian mixture is found by employing maximum likelihood principle with an information criterion, minimum message length principle,for determining an optimal density estimation of the PDF [2]. Restart is conducted by moment-matching sampling the Gaussian mixture, which strictly conserves mass, momentum, and energy. We demonstrate the effectiveness of the method with various electrostatic and electromagnetic particle-in-cell simulations.
[1] Chen and Chacon, “A machine-learning checkpoint/restart algorithm for particle-in-cell simulations”, in preparation
[2] McLachlan and Peel. Finite Mixture Models. John Wiley & Sons, 2004.
Comments:
