High Performance Computing users meet at Rice University

Sandia Labs on bugs, ConocoPhillips’ HPC, ECLIPSE benchmark, GeoBenchmark, Solid State disks.

Curtis Ober presented work done at Sandia National Laboratories on verification of complex codes by exploratory simulation. Ober’s starting point was Les Hatton’s 1997 paper* titled ‘The T Experiments, errors in scientific software.’ This concluded that software accuracy is greatly undermined by software errors and that most packages are full of statistically detectable inconsistencies in programming language use.

Ober’s work investigates the impact of software bugs on numerical error and uncertainty. Numerical error can be broken down into discretization, round-off, conversion errors or ‘implementation coding errors’ i.e. bugs. Defect detection leveraged statistical and other analysis of ‘model code.’ The study found that many bugs can be detected using simple tests. Sandia uses these techniques to validate its own codes

A ConocoPhillips presentation by Bill Menger discussed the delicate balance of Bandwidth, Throughput, IOPS, and FLOPS. According to co-author Dave Glover, a supercomputer is a computer that is one order of magnitude slower than what scientists and engineers need to solve their current problems. This definition is timeless and assures job security!

Menger advocates task driven design, adapting hardware to ensure that users get what we need, no more, no less. One use case called for a 12 million CPU hour run to be completed in 50 days. The solution involved running each job on 8 cores, optimizing data loading, adding memory and disk and ‘buying enough computers to make it happen!’ The chosen hardware included 1250 Rackable Dual Socket, Quad-core AMD Opterons, 10GigE card in each node with iWarp and 16 shelves of Panasas Storage for a theoretical 46 Teraflops.

Gilad Shainer (Mellanox Technologies) presented an investigation of Schlumberger’s Eclipse reservoir simulator performance on network type. A 4 million cell three phase black oil model with 800 wells was used for the test which was run on a 24 node Dell PowerEdge SC 1435 cluster with Quad-Core AMD Opteron 2382 processors. Tests showed that InfiniBand interconnect provided highest performance and scalability up to 8 nodes. Running eight jobs per node increased productivity by up to 142%. Infiniband also reduced power consumption by 82% compared to 10GigE.

Evgeny Kurin of Russian GeoLab introduced a new seismic processing benchmark for HPC. GeoBenchmark includes a set of simple seismic processing modules. Tests target specific computer subsystems and are ‘reasonably portable.’ GeoBenchmark can be used to evaluate the relative contributions from code and compiler. Results are presented as a ‘tuning radar,’ a spider plot showing each subsystem’s contribution to performance. GeoBenchmark source code can be downloaded from www.oilit.com/links/0905_4.

Guillaume Thomas-Collignon (CGGVeritas) presented the results of a series of benchmark tests of new solid state drives from Intel. The Intel X25-E SLC SSD produces spectacular improvement in random read/write producing around 20,000 IOP/S compared to 400 for a high end spinning SATA drive. The downside is a smaller capacity and higher cost per GB. A trace sort test showed a ten fold speed up with the SSD. For Collignon, the Intel X25-E SLC is the way to go. The only current drawback is the low capacity of current drives.
HPC papers available on www.oilit.com/links/0905_6.

* www.oilit.com/links/0905_3.

Click here to comment on this article

Click here to view this article in context on a desktop

© Oil IT Journal - all rights reserved.