Protocol of high speed data transfer from Particle Image Velocimetry system to supercomputer

The paper presents an approach to the transmission of large data flows from a PIV (Particle Image Velocimetry) system to a remote supercomputer via a high-speed optical network. The approach is based on the model of a direct entry of PIV measurements into the RAM of computational nodes of a supercomputer for parallel processing. We describe a network infrastructure designed using the proposed model and a data exchange protocol, which employs any stream mode reliable data transfer protocol (TCP, UDT) on transport level. Excluding the master node and storage of the supercomputer from data processing, we have constructed an algorithm for dispatching parallel data streams to many computational nodes. The dispatch problem is solved on the local computer of the PIV-system using the developed scheduler software. The one- to-one mapping of measurements on computational nodes is abandoned in favor of using an output data queue and the First-In-First-Out (FIFO) strategy for data processing in queue. The possibility of processing the data from a single experiment concurrently on several supercomputers is shown in this paper. Such data processing cannot be realized within the classical approach because it requires a common storage and inter-supercomputers synchronization of processes. The developed approach can be used to perform feedback experiments and to apply computationally complex data processing algorithms that are not accessible on the local computer. The paper also contains examples of using architectural solutions developed for the calculation of the velocity field in the PIV-system and evaluation of the effectiveness of these algorithms.