Results 1 - 1 of 1
"... This research will show that it is possible to provide coupling between components in a simulation using a scalable intercept in the file system I/O interface. This intercept will be shown to be scalable, high performance, provide an easy to use mechanism for coupling components, and which does not ..."
Abstract - Add to MetaCart
This research will show that it is possible to provide coupling between components in a simulation using a scalable intercept in the file system I/O interface. This intercept will be shown to be scalable, high performance, provide an easy to use mechanism for coupling components, and which does not require system level changes for use. The common method for data transference between physics simulations is via the file system. This is a consistent performance bottleneck. Multiple techniques have been proposed to deal with this, e.g. parallel file systems, parallel write threads, parallel I/O, hierarchical I/O, and write combining. But all still have the fundamental flaw of making use of a physical hard drive, which suffers from limited bandwidth and physical limitations on write speed, including disk latency, caching effects, and the limitations of the bus connection to the hard drives. While solid state disks have the potential to alleviate much of this problem, the problems associated with them (such as performance degredation over time and expense) serve to limit their adoption in most applications. This work aims to alleviate this bottleneck in data exchange, thereby tightening the coupling of physics simulations. By making use of direct, live network connections between simulation components, data can be passed at network speed, thereby bypassing the disk bottleneck. By modeling the interface on the extant MPI-I/O interface, it is hoped that scientists and simulation developers can easily deploy this functionality, and make use of it in a fashion to which they are already accustomed. As such, high performance data exchange can be achieved, while maintaining an easy to use interface, similar to extant systems, with minimal changes to existing code bases, and no system wide changes necessary. While the work is specifically aimed at the needs of the Center for Simulation of Wave Interactions with Magnetohydrodynamics (CSWIM) at Oak Ridge National Laboratory, its applicability is much broader, due to the generic I/O interface being presented. As such, many other scientific simulation projects should be able to use it, without a need for specialization. 1 2