Attendees to the Oil and Gas High Performance Computing event held in Rice University last month heard from Intel’s Rajeeb Hazra who summed-up current thinking on the path to ‘exascale’ computing thus: ‘There are those that believe that virtually everything we use today, hardware and software, cannot evolve to work at the exascale. They would have it that a revolution is needed.’ There are many reasons for such a world view, with considerations in both hardware and software field. Except that, for Hazra, these are mostly myths!
Intel is getting HPC back on the path to exascale with the general purpose many core architecture a.k.a. Knights Corner and the Intel MIC. Along with the new hardware comes a scalable ‘ecosystem.’ ‘Today’s architectures and applications can evolve to exascale. A complete revolution is both unnecessary and unaffordable.’
The MIC was the subject of a presentation by Lars Koesterke of the Texas Advanced Computer Center which should be running at 10 petflops next year—’80% down to the MIC.’
Perhaps not revolutionary (Texas Instruments have been making digital signal processors for decades), but certainly different was Murtaza Ali’s presentation on the use of TI’s ‘Shannon’ multi-core DSPs for seismic imaging. The Shannon is based on the KeyStone multicore architecture integrated with eight C66x CorePac DSPs per chip. A 1 teraflop card from Advantech should be available real soon now. Ali is working with Rice University’s Jan Odegard to implement ‘Iwave,’ a framework for scalable finite difference seismic modeling. Early results are promising.
Another Iwave port was presented by AMD’s Ted Barragy, who teamed with Rice’s Bill Symes to translate Iwave into OpenCL and run it on AMD’s ‘Fusion’ processor. According to AMD, the GPU is passé. What’s hot is combo CPU/GPU technology that is taking off ‘exponentially.’ Such a beast is AMD’s Fusion, combining the CPU and GPU into a single device. Will Fusion be good for seismic processing? Yes according to Barragy, at least when the ‘full’ Fusion arrives. On the coding front, the OpenCL port went well thanks to Iwave’s clean design and the OpenCL helper libraries and the existence of a strong tablet to desktop market to commoditize the technology.
University of Houston researcher Deepak Eachempa, with backing from Total, presented an evaluation of Co-array Fortran (CAF). CAF is intended to simplify porting of the substantial Fortran code base to modern parallel compilers. The idea is to effect the smallest change required to make Fortran an effective parallel language. Early results show CAF to beat IntelMPI and OPenMPI. CAF is ‘very promising as a programming model for oil and gas HPC applications.’ CAF is now available in the OpenUH compiler which can be downloaded.
Jeff Davis gave a more down to earth presentation of technical computing in the new Hess tower. This has gained ‘leadership in energy and environmental design’ (LEED) certification for its general greenness. The Tower contains 944 miles of Fiber optic cabling and ‘even the window shades have IP addresses.’ Hess’ HPC cluster includes 2,256 Nvidia GPUs, 5,160 CPU cores and a NetApp client server system for interpretation. Geoscience workstations come with 192GB Ram, 12 cores, 10 Gig networking, SSD disks and more. The Tower also houses 50 high end visualization rooms linked with a connection broker supporting both Microsoft and RedHat operating systems. The latter was developed with Mechdyne using HP’s Remote Graphics. More from O&G HPC at Rice.
© Oil IT Journal - all rights reserved.