Automating computational experiments data post-processing
-
D.V. Leontiev,IAC PFEBRAS Vladivostok. Russia
-
Kharitonov D.I.
D.I. Kharitonov. IAC PFEBRAS Vladivostok. Russia
-
Odyakova D.S.
D.S. Odyakova. IAC PFEBRAS. Vladivostok. Russia
-
Parahin R.V.
R.V. Parakhin IAC PFEBRAS. Vladivostok. Russia
This article considers the principles of the event control system for processing data from computational experiments. An approach to construct a data processing models of a computational experiments is considered. To make models the Petri nets are used. The
model of computational experiment consists of computational and control processes models. The models are built separately. The computational process model is built in a two stages. On the first stage the user generates the event tree of computational ex-
periment. On the second stage the computational process model is automatically built from the event tree. The model of the control process is built from a reaction patterns. The following three reaction patterns are developed: reaction on a previous event, reac-
tion on each N-th event, reaction on a next event. The reaction pattern is configured on the triggered event. The approach allows users with minimal skills to make the data processing models. The architecture of the event control subsystem of a computational
experiment is considered. A description of the tools used (Slurm and Audit), which are the basis for the functioning of the event management system, is given. Event control is performed using the Audit subsystem, which collects events and sends them to the
processing node. A control process is located on the processing node, which track events on the model and starts the execution of the corresponding reactions. A description of the starting process computational experiment with event control is given. The sequence
of events convolution algorithm is described, which is designed to search for inconsistencies between model and a real computational experiment. The main feature of the developed approach is that there is no need to reprogram an original computational task.
Keywords: high performance computing, big data processing, multiprocessor computing systems, scientific data visualization, Petri nets.