When looking at processors today, Barry notes that basically any type is OK, be it Itanium, Opteron, Power, Sparc, (but not T1) provided it is 64 bit. Because the processor cost is not a big part of the overall system cost, one looks for the fastest available. You can look at available bench marks for that, but you have to be aware. For instance PC performance reported today in the world is Dry Ice cooled Opteron processor, perhaps not something you want to have in a full system run in a data centre.

One way to accelerate the performance further for reasonable power cost is specialized co-processors. At Airbus they are looking for instance at

• VAP: vector attached processor using SIMB style vector processing
• FPGA: putting the application in hardware
• PIM: processor in memory.

According to Barry, PIMS could well be a dead-end: there are not yet any useful applications that are using it.

To effectively use a vector processor the code needs to be optimized: "vectorized". Barry showed that even for other architectures the vectorized Airbus application showed a speed-up. on the different processor types the figures are:

EPIC 20x

RISC 10x

CISC: 2x

which clearly shows vetorizing code is not a dead art.

Another problem is storage and access of data. Today RAID6 provides 99.999% accuracy. To do better, for instance with RAID11, one needs to double the disk storage which double the power consumption. Disks - the databases that are on the disks - are on-line always. One does not shut down disks in a data centre. So this becomes a major management concern in the data centre.

Another concern is the increasing density of the systems. Data centres were built for 1000KG/rack weight, 4KW/rack power consumption. Today's blades and disk arrays use already more. How to handle that? How do you get enough power in your centre and how can you dsitribute it? At Airbus they are, for instance looking at fans for racks instead of fans for servers; and look at diskless servers.

HPC is now pervasive in Airbus. It is a core technology for all design. The HPC systems must work correctly and provide timely results. Airbus requirements on HPC are 100% availability, guaranteed performance, support a long application lifetime ("101 years"), lower system power consumption and a smaller system footprint.

The HPC user community at Airbus grew with its importance: from about 50 in 1990 to 5000 today. As a result today's HPC users are really users, as opposed to the 50 that were HPC experts. The experts could handle and manage the supercomputers on their own. Today that is impossible. Airbus depends on HPC for design optimisation and analysis; and to support manufacturing and services too much.