In 1985, we witnessed the birth of Cray-2, a system with 4 processors, 244 Mhz, 1,95 Gflop/s peak, 2 GB of memory, 1.2 GFlop/s linpack R_max, 1.6 square feet of floor space and 0.2 MW power. In 2005, the BlueGene/L system is the number one in the TOP500 list with 700 Mhz, 65,536 nodes, 180 Tflop/s peak, 32 TB of memory,

135 Tflop/s Linpack R_max, and 250 square feet of floor space, showed Horst Simon.

Between 1985 and 2005, we saw custom built vector mainframes evolve towards commodity massively parallel platforms. In 1985, 30 Mflops sustainded was good performance but now there is already 1 Tflops sustained necessary to make good performance. In 1985, vector Fortran was used to program versus Fortran/C with MPI at this instant.

Dr. Simon raised the suspense in the audience while running through the personal list of 10 resans why supercomputing dramatically changed between 1985 and 2005 that he drew up for this anniversary conference.

Simon's number 10 is the TOP500 list, at 9 figures the NAS parallel benchmark, at 8 the Grid, at 7 the hierarchical algorithms including multigrid and fast multipole methods, and at 6 the HPCC initiative and Grand Challenge application from a United States point of view.

At 5, he placed the "Attack of the killer micros", a term first coined by Eugene Brooks (LLNL) at Supercomputing 89. It became a catchy shorthand expression for the technology change from custom ECL to commodity CMOS.

Commodity CMOS micro processors did change the face of supercomputing but they were neither inevitable, nor the only technology choice, according to Dr. Simon.

At 4, we find the Beowulf clusters. Thomas Sterling et al. established a vision of low cost, high-end comuting which demonstrated the effectiveness of PC clusters for some but not all classes of applications. This team provided software and conveyed findings to the broad community through tutorials and a book in 1999.

The approach made parallel computing accessible to a large community worldwide and also broadened and democratized HPC. However, it effectively stopped HPC architecture innovation for at least a decade.

At 3, Horst Simon surprisingly market scientific visualization. The NSF report "Visualization in Scientific Computing", edited by B.H. McCromick, T.A DeFanti and M.D. Brown, established the field in 1987. It market a change in point of view and transformed computer graphics from a technology driven subfield of computer science into a medium for communication. It also added an artistic element

At 2, there is the Message Passing Interface, called the default standard by Horst Simon but he admitted that it was a very difficult transition. At the 1988 Salishan conference there was a take-off of parallel programming languages trying to solve five scientific problems. The Salishan Problems invesigated four programming languages: Sisal, Haskel, Unity, and LGDF. It formed a significant research activity at the time and the early work on parallel languages is all but forgotten today.

Parallel language programming existed in the academic world but not in the industry. The availability of real machines moved the discussion from the domain of theoretical CS to the pragmatic application area. In 1994, there was an evolution in programming models. Horst Simon stated that in 1988 we had different architectures and a single programming model; in 1993 there was a single architecture but multiple programming models; and

in 1997 there were multiple architectures with multiple programming models. At NERSC in 1996 among scientific users 30% used PVN and 30% used MPI.

Horst Simon asked himself in 2001 what would be the new programming modality by 2010. The software challenge consists in overcoming the MPI barrier. MPI finally created a standard for applications development in the HPC community. Standards are always a barrier to further development, according to Dr. Simon.

At 1, Horst Simon surprised the audience with scaled speed-up. We should grow for this problem as a conceptual breakthrough, he stated. In the late eighties, MPP were due to fail, one thought because of the speed-up argument. The argument against parallelism in 1988 was that one would need infinite parallelism. This held back the building of MPPs during several years. It helped however the community to overcome the conceptual barrier

and exposed the fallacy of the fixed size speed-up.

As for the future challenges, Horst Simon divided the coming decades into three periods:

• 2005-2012
• 2010-2018
• 2015-2025

For 2005-2012, there are three major issues: scaling to 100.000 processors and multi-core processors, a topology sensitive interconnection network, and the memory wall. The norm for average applications now is a few Teraflop/s of sustained performance, scaled to 512-1024 processors. The number of processors in most highly parallel systems has not changed in eight years between 1997-2005. Now this has changed with BlueGene/L.

Scaling the number of processors is difficult because of inefficiencies of MPI, thus we face programming language barriers. The tools are not scaling so there is work to be done, urged Dr. Simon.

The challenge for 2010-2018 is to create a new ecosystem for HPC. Platforms, software, institutions, applications and people who solve supercomputing applications can be thought of collectively as an ecosytem. Research investment in HPC should be informed by the ecosystem point of view: progress must come on a broad front, according to the speaker.

In the 1980s Cold War and Big Oil spending were the goals to build vector supercomputers: they are the dinosaurs in the ecosystem. We also saw 20 years of Fortran applications based in physics codes and third party applications as well as commercial off the shelf technolgoy (COTS). Further more, clusters emerged and there were 12 years of legacy MPI applications. There also has been a massive R&D investment.

HPCC in the United States has witnessed a vigorous computer science experimentation in the languages tools system software and the development of the Grand Challenge applications. An external driver has been the industry transition to CMOS micros. All changes happened vitually at once so the supercomputer community faced an ecosystem change, according to Dr. Simon.

In 2005 the ecosystem is very stable. There has been an attempt of re-introducing old species such as the X1 but this failed but there also have been attempts of introducing new species and Horst Simon cited the mutation of Blue Gene 1999 to BG/L in 2005. It seems to work well, stated Horst Simon: just look around the room.

So why is not everybody happy and content?

The supercomuting community is aware that the ecosystem is far from optimal because of the divergence problem and Blue Planet at NERSC. There is also a major concern about the right benchmarks. The current ecosystem will become untenable after about 2010 in the face of the architectual and software challenges, warned Dr. Simon.

How are we going to change the ecosystem? And what are we going to change it into? According to Dr. Simon, DARPA HPCS is on the right track combining requirements for new architecture, new languages and insistence on commercialization by vendors. But will a 150 million dollar programme be enough to change a 6-8 million dollar market?

As for the 2015-2025 challenge, some time between 2015 and 2025 the continued performance growth of semiconductor based microprocessors will end. Dr. Simon cited the Vanishing Electron element referring to the transistors on a chip and their number.

Dr. Simon compared the End Of Moore's Law to FM Radio. When driving away from the FM transmitter, you get less signal noise from the electrons. There is no change. Similarly, when you increase the numbers of gates, you get less signal power noise from the electrons. Again, there is no change. This cannot go on forever.

As for the 2013-2016 period, many solutions are not known today. For the 2015-2025 challenge, there is no active, strong research programme anywhere that addresses this challenge. Alternative technology solutions are feasible but won't come by themselves. After 50 years of exponential growth, how will the industry adjust to a no-growth scenario, asked Dr. Simon.

Therefore it will be necessary to learn how to scale, to change the ecosystem, and to deal with the flattening of Moore's Law, Horst Simon ended his talk.