Keith Gray stated that the HPC has expanded form 0.1 Tflops to 10.2 Tflops in the periode from 1999 to 2003, thus with a factor of over 100 percent. BP has reduced cycle time for a Thunder Horse migration from 3 weeks to one day. The speaker provided the audience with examples of seismic migration and multiple attenuation, with a simple example of wave propagation, and computer visualisations of seismic sound waves.

The Thunder Horse is located about 150 miles southeast of New Orleans in over 6000 ft of water. CGG Symphony acquired a proprietary seismic survey over the Thunder Horse discovery in the deep water Mississippi Canyon area, as Keith Gray mentioned.

HPC skipped a significant investment in IA-32 clusters. Programming difficulties would not optimize the imaging research effort.

Offering an overview of the project history, Keith Gray told the audience that first Linux clusters were used in HPC early 2000, called Arco clusters. Later the testing started with Itanium. In September 2001 BP worked very closely with Intel. In September 2002 there was a pilot test with 15 HP rx5670 servers and in

November 2002, 131 HP servers were deployed.

The applications supported are seismic migration codes in the internally developed seismic research at BP. In the elapsed time dominated by computers, I/O is not as critical but I/O is more key than migration and will require more capable clustered file systems.

Application infrastructure consists of Fortran and some C. There is mixed mode parallelism MPI and OpenMP with complex multiple 1DFFT, complex 2DFFT, complex 3DFTT, and real to complex FFT. The common I/O interface is in C. There is a data dictionary for exchange of multiple formats, an Intel compiler and an HP Math library.

HPC computing architecture is built out of the computer systems today: an HP Itanium 2 computer cluster with 395 nodes. Each node is 4 x 1.3 Ghz Itanium 2, and has 32 GB memory. There are five SGI Altix systems.

The storage delivery is done by SGI CXFS clusters and SGI NFS servers for a direct attached storage. There is 300 TB of total RAID storage, LSI Metastor E3300 RAID - used sinde 1998, Infortrend RAID - in use since April 2003, and

SGI TP9300. The move to serial ATA was driven by cost, as Keith Gray explained.

The goal is to meet the I/O requirement of the Itanium 2 cluster and to be able to deploy as many file systems as required, cost-effectively. The speaker talked about over 1 GB/sec Read performance and of over 20TB file systems. There are SGI Origin 300s as big NAS servers. The storage growth forecast for 2005 will maybe amount to over 600 TB. At present, it is 300 TB RAID storage, in 1999, it was 12 TB RAID storage.

Keith Gray also addressed the tape requirements. There are two reasons for tape, namely the off-site archive and the need to reduce disk demand. The systems today in use are Grau robots and AMpex DD-2 tape drives. Future requirements include common interchange media with geophysical acquisition. For the testing of LTO-2, BP will watch IBM 3592. BP makes no back-up, there is no window but it can recreate data. BP will manually use tiered disk if it is cost-effective. The

goal, according to Keith Gray, is to have enough disk so active projects do not move to tape.

The speaker also mentioned the HPC computing cost decline between 1997-2005. He concluded by asking what would be the limit if growth is beyond the current rate? People and computer room space would be issues to face. Therefore, BP has to work to meet smart people to have options. The Horsespace issue can get solved with money.