HPC PROJECTS: OIL AND GAS
➤
to seismic processing. SRC Computing, based in Colorado Springs, USA, develops computing systems based on reconfigurable processors – accelerator boards consisting of arrays of FPGAs. ‘There’s been something of a love affair between the oil and gas industry and GPUs, and this is because for a number of years the holy grail of the industry has been visualising large chunks of the data – that need was delivered by graphics cards,’ says Mark Tellez, director of business development at SRC. Tellez believes that the move from visualisation applications to GPU-led processing was something of a natural progression, as many client companies have already invested in the necessary hardware. ‘In a lot of ways, GPUs were just a product looking for a solution that they can solve, but [GPU suppliers] were not building a compute solution, they were building a graphics card designed to solve a graphics problem,’ he says.
SRC supplies customers in the oil and
gas industry with a way of speeding up the demanding reverse time migration algorithms used in processing seismic data. David Caliga, director of software applications at SRC, describes the company’s approach: ‘We provide what a lot of people might call accelerators, capable of speeding up compute-intensive portions of code. What we provide is a complete system, and not just a GPU- or FPGA-based accelerator card.’ The company’s products are based on highly parallel and reconfigurable FPGA accelerators, which it refers to as MAPs. Caliga states that the company optimises its systems to ensure that seismic data can be very rapidly moved into the MAPs. Several MAPs may be incorporated into a single system, working alongside the standard CPU microprocessor. ‘We treat the MAP processor as a peer to the microprocessor. The intent is to divide the compute intensive application across the two compute devices to get the most out of both of them. In one image processing application, for example, a system with five MAPs and one microprocessor replaced a cluster with 96 dual-core nodes.’ The performance of MAPs-based systems is described by the company as comparable to that of GPU-based systems while only requiring around a quarter of the power
26
of an equivalent GPU system. ‘We can provide superior performance with a fraction of the power dissipation,’ states Caliga. ‘Fully loaded, the MAP consumes approximately 50W, compared to a GPU, which would be around 200W.’ A reduction in power consumption and heat dissipation
‘In a lot of ways, GPUs were just a product looking for a solution that they can solve’
also allows a reduction in the footprint of the system: ‘The potential reduction is from something that would consume multiple racks in a data centre down to something that would fill a small desk-side enclosure, or maybe a standard rack.’ Tellez explains that even in a multi- billion dollar industry, power consumption is of key importance: ‘I know of a number of [seismic data processing] companies that are constrained by the amount of power they can pull from the grid at the location of their data centre. Additionally, even if they can pull enough power for their server farms, they’re typically working in a building that was never designed to have that much power generated on site (in terms of heat). They therefore have trouble in terms of cooling.’ He adds that many companies are now looking towards moving data processing operations nearer to the drilling location, on a ship, or even
on the drilling platform itself, where power and space are even more constrained. ‘Typically, they put the data onto a number of hard drives and fly it by helicopter back to the mainland, where it is loaded into a computer. They compute whatever they are working on, and then they often have to send the results back in the same manner! If they could do all of the processing on site, they’d be saving a lot of time and money, not only in terms of just moving the data, but also in terms of avoiding idle time on the drilling platform while they wait for results.’ SRC believes therefore that its low-power processors offer the oil and gas industry an attractive alternative to both conventional clusters and GPU-based computation. Tellez states that SRC has gone after customers in the industry of seismic data processing because importance of processing in the industry: ‘They are the low-hanging fruit at the moment, because the faster they can process a line of code, the faster they can either begin drilling or charge the drilling company for the information.’ Given the advantages of FPGA-based
processing over cluster or GPU-based alternatives, why haven’t more oil and gas companies adopted the company’s solutions? Tellez believes that FPGAs (or reconfigurable processors) are seen as complicated and difficult to program. ‘The biggest challenge we have is to get people to realise what they can do with the current technology, and to realise that it’s not as scary as they may have heard.’ In a move analogous to Nvidia’s introduction of the Cuda programming environment for GPUs, SRC offers its CARTE platform for programming FPGAs. CARTE contains a set of standardised functions specific to the oil and gas industry, and the company hopes that this will enable more and more players in the industry to take advantage of the reconfigurable technology. ‘We’ve done a lot of work in order to simplify the move from the original calculation into the reconfigurable environment,’ says Tellez. Before the introduction of GPU programming environments such as Cuda and OpenCL, GPUs were very difficult to program. As programming tools for reconfigurable environments become more established, programming the right tool for the job can only become easier.
SCIENTIFIC COMPUTING WORLD JUNE/JULY 2010
www.scientific-computing.com
C M J CM MJ CJ CMJ N
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55 |
Page 56 |
Page 57 |
Page 58 |
Page 59 |
Page 60