Conventional wisdom has it that reservoir flow simulation is hard to parallelize. Most simulators are limited to around a few tens of cores. Rock Flow Dynamics claims to have broken the core count barrier in a test of its tNavigator simulator running on Moscow State University ‘Lomonosov’ cluster. The trial data set was a real field black oil model with 43 million active grid blocks and 14,000 wells. RFD’s approach centers on the use of compression to speed data access. Input data files are compressed 10x with the public domain Gzip library. To demonstrate the use of potentially sensitive data in the cloud, data was also encrypted. The model ran on an increasing number of nodes of dual four-core Intel Xeon 5570 Nehalem processors—up to 512 nodes (4096 cores). Speed-up maxed-out at 350x on 2048 cores.
In another test, on a 22 million cell three-phase model, calculation time was reduced from 2.5 weeks down to 19 minutes after numerical experiments and tuning. This represents a 1300x speed-up over a single core which is probably, according to RFD, a world record. The test showed linear speed-up on up to 2048 cores. From then on, core count doubling to 4096 showed a slight tailing off—but still produced a 1.4x boost.
RFD has also performed tests comparing Nvidia Tesla and ATI Fire GPUs against the Xeon and have found that re-coding to CUDA and OpenCL is ‘far from trivial.’ Moreover, the tests show that ‘benefits from the CPU/GPU combo remain unclear—and it may take generations of CPU architecture to become competitive with the Xeon.’ Read the full test results in SPE paper 163090. More from tom.robinson@rfdyn.com.
© Oil IT Journal - all rights reserved.