Michael Wolf (Sandia National Labs) described how to achieve ‘painless parallelism’ on multi core architectures with the ‘Trilinos’ object-oriented framework for large scale science. Trilinos lets scientists write once and run on a variety of shared memory architectures including multi core CPUs, GPU and NUMA. More from trilinos.sandia.gov.
Vlad Bashkardin (UT Austin) was circumspect on the often reported ‘orders of magnitude’ performance hike of GPUs. These frequently compare optimized GPU code against an outdated CPU. Using reverse time migration and test data supplied by Chevron, UT’s Texas Advanced Computer Center reports a more modest 10 to 15 fold speedup for RTM code—itself something of a best case for parallelization. For more sophisticated imaging 10 x and 200 Gflops would be ‘excellent.’
Jens Kruger outlined the Fraunhofer Institute’s ‘Green Wave’ Project addressing power hungry HPC tasks like RTM.
Green Wave leverages Tesilica’s customizable processor cores as deployed in the Berkeley Labs Green Flash climate modeling project. Tensilica allows for the development of a processor that is optimized for a particular application. Before the chip is actually fabricated, the Berkeley Emulation Engine* is used to predict performance upfront. GW falls between the Nehalem and Tesla in terms of performance but wins out in terms of ‘megapoints’ per watt. The first actual chip will come off the fab line in Q4 2011.
Maxeler’s Oliver Pell also stood out from the GPU crowd advocating RTM on an FPGA—with another 1000 x speedup claimed. FPGAs offer terabytes of memory bandwidth—but only a 100MHz clock. Maxeler has bundled its FPGAs with a high speed ‘MaxRing’ interconnect—with MaxBoxes surrounding an X86 CPU controller. The MaxGenFD Java library promises good scalability. ‘Propagating 70Hz waves is practicable.’
Gladys Gonzalez described Repsol’s search for next generation computers and programming tools. GPU accelerators outperform the CPU by 20x but there are serious issues with host to accelerator interoperability, with code portability and maintenance. Industrial software needs industrial standards. Today programmers need to know every detail of the stack. There is good news coming from the Barcelona center for Super Computing’s GMAC** library offering a unified virtual address space and data management. OpenCL also got a plug as a potential standard replacement for NVIDIA’s CUDA.
© Oil IT Journal - all rights reserved.