Features:

EXAMINING SGI'S VISUAL AREA NETWORKING WITH CINES

HPCwire recently interviewed Thierry Porcher, CINES Director, and Olivier Lernout, CINES communication director, regarding SGI's Visual Area Networking and its involvement with CINES.

HPCwire: What, exactly, is SGI's Visual Area Networking (VAN) technology, and how is it practically implemented at CINES?

SGI's Visual Area Networking (VAN) technology involves supercomputers - such as SGI's Onyx2 - and their ability to increase efficiency by allowing creators and consumers to remain in one location and visually interact with large data sets, applications and colleagues located at another site. CINES has 2 supercomputers: A cluster of SMP from IBM (29 x16 way Power3 nodes + 2 x 32 Power4 nodes) and an SGI Origin 3800 (768 processors). Code outputs from these machines are analyzed on an SGI graphics server (Onyx2, 8 CPUs, 4 pipes IR3). Users are able to share data between the Origin 3800 and the Onyx2 thanks to CXFS. French academic researchers are then displaying post-processing results (using Ensight from CEI for example) remotely through RENATER, the National Network for Research. For high performance, they use VizServer on the Onyx and a client application on their workstation.

HPCwire: Please give an example of how a specific complex problem can be solved via VAN technology? How was such a problem attacked prior to VAN, and with what results?

The continual increase in processing power available at national computing centers allows the creation of more complex numerical simulations and more accurate results. However, increases in complexity and accuracy can lead to huge result files that are difficult to exploit. The management of many large pre and post processing files has become a major inhibitor that slows the development of more realistic numerical simulations. Not only has it become costly to transfer files between a laboratory and the computing facility, but the visualization systems required to study such large volumes of data are too expensive for most research laboratories.

"Extreme computing for turbulent combustion", The CERFACS research project, lead by Yannick Sommerer, illustrates these facts. Coupling multi-physics is one of the current areas of research for turbulent combustion where thermal- acoustic phenomena and fluid-structure interactions play a fundamental role in flow stability within combustion chambers. In order to correctly predict the level of noise and pollution generated by turbines, we must simulate large- scale problems on high-resolution meshes and take into account as much as possible the flow physics both before and after the combustion chamber.

Though the CINES computers can process a numerical simulation with a 107 point mesh in a reasonable time (which is already a step beyond traditional capabilities in this domain), the post-processing of the generated data is a computing challenge on its own.

• Such a large simulation cannot be done in one step. The computation must be regularly checked to make certain that it is converging properly. Intermediate results are regularly post-processed to validate the results with about ten 80MB files required for each iteration. This results in a transfer of over one half a gigabyte of data for each intermediate check.
• At the end of the simulation, over one hundred result files of 80MB each (the result of ten intermediate phases) files have to be processed through a powerful visualization server.

With a sustained 400Kb/s network bandwidth between CERFACS and CINES, data download was the main bottleneck for this numerical simulation. The download time would have been about 4.5 hours for each post-processing check, and about 45 hours for the final data set to be exploited!

Following the event "Visualization and Large Scale Computing" organized by CINES on October 21st 2002, we were interested in evaluating the features of OpenGL Vizserver computing solution and its ability to solve our problem.

In summary, visualization was done on the Silicon Graphics Onyx2 visualization system located at CINES, and the graphical results were sent across the network and with OpenGL Vizserver, remotely displayed on a client located at CERFACS, more than 200 km away. In this way, post-processing can be done through the network without transferring any data sets except images which can be compressed using different techniques to fit the available network bandwidth.

HPCwire: What will CINES use VAN for in the upcoming years? Will VAN upgrades be necessary?

Added or new functionalities would be nice : pipe sharing between multiple users, better collaborative sessions ...

CINES intend to increase its computing power to reach at least 10 TFlops beginning of next year. We expect a large increase of the amount of data generated by simulation codes and more powerful graphics server will be needed.

HPCwire: How does the cost of VAN compare with competing technologies?

Other systems would require a larger network bandwidth or to use more sophisticated network mechanism (QoS). Using VAN allows us to cope with the current network infrastructure while being very easy to manage.

HPCwire: How easy is VAN to use? How much special training is necessary? How scalable is VAN?

Very easy ! Installing the client on your PC or workstation take 5 minutes and using Vizserver is straightforward.

HPCwire: Are there any other things you'd like our readers to know?

Not yet !

ABOUT CINES

CINES is a world renowned supercomputing center that provides high-end resources to R&D facilities located throughout France. CINES users routinely generate multi-gigabyte files containing CFD, Structural and Dynamics simulation results that they then must analyze. The typical method of doing that analysis was to transfer all of the results to local workstations and then hope that those workstations have enough compute, memory, I/O and visualization capabilities to enable thorough analysis. Because many of their users lacked desktop resources to properly analyze results and because data transfer times often ran into many hours or days - decreasing the rate of innovation and discovery. Furthermore, with planned increases in HPC and large data resources at CINES, these bottlenecks were about to get much worse. CINES wanted to maximize the value obtained from HPC resources, and their customers wanted to accelerate their R&D efforts so they turned to SGI's Visual Area Networking (VAN) technologies to accelerate the analysis process to keep pace with their increasing compute power. Engineers at CERFACS, which is located 250 miles away from CINES, have shown how VAN has decreased their analysis time by a factor of between 3 and 10, enabling them to create more accurate designs in less time than ever before.