The changing faces of supercomputing in the pervasive world of IP technologiesHeidelberg 22 jun 2001 Dr. Horst Simon, Director at the National Energy Research Scientific Computing (NERSC) at the Lawrence Berkeley National Laboratory, in his keynote talk at SC 2001, shared some of his viewpoints with the audience on what he expected will be the major trends for the next five years in scientific computing. These developments include the rapid growth of micro-processors, the increasing use of open software; the optimised network bandwidth which will grow at an even faster rate than Moorḙ墳 Law has predicted; the somewhat paradoxical evolution towards aggregation, centralisation, and co-location; and the emergence of commodity products which will trigger the use of IP on everything that you can imagine.
The speaker announced the appearance of the first teraflop system by 2005. This is about one to two years later than planned in the ASCI scenario. From then on, no system with less performance than one Tflop/s will make it to the TOP500. The first petaflop system will see the day around 2009 which calls for rapid but unpredictable changes in HPC technologies. In this landscape, clusters are omnipresent with a 10 Tflop/s at Lawrence Livermore, a 3 Tflop/s at NERSC, and a 30 Tflop/s Compaq system at Los Alamos National Laboratory. In the United States, there is a preference for SMP nodes whereas Japan sticks to vector nodes but basically these two types of supercomputing approaches are nothing less than large cluster machines. The Japanese Earth Simulator is an example of a cluster machine.
IBM's Blue Gene can be considered as a petaflop system "avant la lettre". Dr. Simon thinks it to be a great system but with still too much limitations for real petaflop problem solving. In turn, the contributions of Beowulf form an interesting experiment in parallel computing systems which demonstrate a great effectiveness of PC clusters for specific classes of applications. MPI has been set as the standard for global communication but according to Dr. Simoṋ墳 belief, standardisation of MPI could mean a barrier to future development. MPI is in fact mid eighties technology and old history, but it still serves as the current programming model at Los Alamos and NERSC.
Open source software as promoted by Linux is definitely the future, according to the speaker. Particularly for high performance computing, open software is important because the HPC community is small. In addition, without this possibility, a lot of experimental but promising programming is threatened in its further existence. Open software thus can prevent the loss of intellectual investment.
Next to this, the scaling of supercomputers will expand the possibilities for tele-collaboration. Dr. Simon described the emergence of two schools of thought, being Internet computing and grid computing. An example of Internet computing is the by now familiar SETI@home project, in which private PC users offer unoccupied PC processing time to an initiative to search for extraterrestrial signs of life. No matter how fascinating this type of supercomputing may be, the speaker has more faith in grid computing which provides middleware to interconnect resources. Grid activities will provide an integration of data, computing, and experimental resources but higher bandwidth will not necessarily lead to distributed computing, however.
The newest hype is to implement IP on all sorts of domestic appliance devices with general accessibility through intuitive interfaces. In the States, a conference was recently organised around the topic of supercomputing with wireless handheld devices. Non-PC devices will be connected to the Internet by means of system-on-a-chip technology in contrast with the most powerful supercomputers which still will increase in size. Dr. Simon in this regard showed some examples of large concrete buildings to host the machines. In the past years, he has been more of a construction specialist than a computer scientist.
As a gadget for the audience, Dr. Simon held in store a science fiction (?) tale about Microsoft. By 2004, this company would be split up into three parts. The application division would then acquire the near bankrupt Cray Inc. Dollars would be spent in re-boosting the Tera MTA while Microsoft would lose its grip on the proprietary key applications such as word processing, spreadsheets, and so on which would then be provided by open source competitors. Or in other words, a story that repeats itself but differently. Has Cray indeed not lost sight of visualisation opportunities?
Leslie Versweyveld |
 |